Added training configuration (fairseq)
Browse files
README.md
CHANGED
@@ -1 +1,47 @@
|
|
1 |
-
Wav2vec2.0-xlsr-53 model is fine-tuned on commonvoice russian dataset
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Wav2vec2.0-xlsr-53 model is fine-tuned on commonvoice russian dataset
|
2 |
+
|
3 |
+
Configs (yaml)
|
4 |
+
-checkpoint
|
5 |
+
---save_interval: 1000
|
6 |
+
---save_interval_updates: 1000
|
7 |
+
---keep_interval_updates: 1
|
8 |
+
---no_epoch_ckechpoints: true
|
9 |
+
---best_checkpoint_metric: wer
|
10 |
+
-task:
|
11 |
+
---_name: audio_finetuning
|
12 |
+
---normalize: true
|
13 |
+
---labels: phn
|
14 |
+
-dataset:
|
15 |
+
---num_workers: 6
|
16 |
+
---max_tokens: 800000
|
17 |
+
---skip_invalid_size_inputs_valid_test: true
|
18 |
+
---valid_subset: valie
|
19 |
+
-distributed_training:
|
20 |
+
---ddp_backend: legacy_ddp
|
21 |
+
---distributed_world_size: 4
|
22 |
+
-criterion:
|
23 |
+
---_name: ctc
|
24 |
+
---zero_infinity: true
|
25 |
+
-optimization:
|
26 |
+
---max_update: 25000
|
27 |
+
---lr: [0.00001]
|
28 |
+
---sentence_avg: true
|
29 |
+
---update_freq: [4]
|
30 |
+
-optimizer:
|
31 |
+
---_name: adam
|
32 |
+
---adam_betas: (0.9, 0.98)
|
33 |
+
---adam_eps: 1e-8
|
34 |
+
-lr_scheduler:
|
35 |
+
---_name: tri_stage
|
36 |
+
---phase_ratio: [0.1, 0.4, 0.5]
|
37 |
+
---final_lr_scale: 0.05
|
38 |
+
-model:
|
39 |
+
---_name: wav2vec_ctc
|
40 |
+
---apply_mask: true
|
41 |
+
---mask_prob: 0.5
|
42 |
+
---mask_channel_prob: 0.1
|
43 |
+
---mask_channel_length: 64
|
44 |
+
---layerdrop: 0.1
|
45 |
+
---activation_dropout: 0.1
|
46 |
+
---feature_grad_mult: 0.0
|
47 |
+
---freeze_finetune_updates: 0
|