| wandb_group: lvdm | |
| output_dir: experiments/lvdm | |
| pretrained_model_name_or_path: null | |
| vae_path: models/vae | |
| globals: | |
| target_fps: 32 | |
| target_nframes: 64 | |
| outputs: | |
| - video | |
| - lvef | |
| - image | |
| datasets: | |
| - name: Latent | |
| active: true | |
| params: | |
| root: data/latents/dynamic | |
| target_fps: ${globals.target_fps} | |
| target_nframes: ${globals.target_nframes} | |
| target_resolution: 14 | |
| outputs: ${globals.outputs} | |
| - name: Latent | |
| active: true | |
| params: | |
| root: data/latents/ped_a4c | |
| target_fps: ${globals.target_fps} | |
| target_nframes: ${globals.target_nframes} | |
| target_resolution: 14 | |
| outputs: ${globals.outputs} | |
| - name: Latent | |
| active: true | |
| params: | |
| root: data/latents/ped_psax | |
| target_fps: ${globals.target_fps} | |
| target_nframes: ${globals.target_nframes} | |
| target_resolution: 14 | |
| outputs: ${globals.outputs} | |
| unet: | |
| _class_name: UNetSpatioTemporalConditionModel | |
| addition_time_embed_dim: 1 | |
| block_out_channels: | |
| - 128 | |
| - 256 | |
| - 256 | |
| - 512 | |
| cross_attention_dim: 1 | |
| down_block_types: | |
| - CrossAttnDownBlockSpatioTemporal | |
| - CrossAttnDownBlockSpatioTemporal | |
| - CrossAttnDownBlockSpatioTemporal | |
| - DownBlockSpatioTemporal | |
| in_channels: 8 | |
| layers_per_block: 2 | |
| num_attention_heads: | |
| - 8 | |
| - 16 | |
| - 16 | |
| - 32 | |
| num_frames: ${globals.target_nframes} | |
| out_channels: 4 | |
| projection_class_embeddings_input_dim: 1 | |
| sample_size: 14 | |
| transformer_layers_per_block: 1 | |
| up_block_types: | |
| - UpBlockSpatioTemporal | |
| - CrossAttnUpBlockSpatioTemporal | |
| - CrossAttnUpBlockSpatioTemporal | |
| - CrossAttnUpBlockSpatioTemporal | |
| noise_scheduler: | |
| _class_name: DDPMScheduler | |
| num_train_timesteps: 1000 | |
| beta_start: 0.0001 | |
| beta_end: 0.02 | |
| beta_schedule: linear | |
| variance_type: fixed_small | |
| clip_sample: true | |
| clip_sample_range: 4.0 | |
| prediction_type: v_prediction | |
| thresholding: false | |
| dynamic_thresholding_ratio: 0.995 | |
| sample_max_value: 1.0 | |
| timestep_spacing: leading | |
| steps_offset: 0 | |
| train_batch_size: 16 | |
| dataloader_num_workers: 16 | |
| max_train_steps: 500000 | |
| learning_rate: 0.0001 | |
| lr_warmup_steps: 500 | |
| scale_lr: false | |
| lr_scheduler: constant | |
| use_8bit_adam: false | |
| gradient_accumulation_steps: 1 | |
| noise_offset: 0.1 | |
| drop_conditionning: 0.1 | |
| gradient_checkpointing: false | |
| use_ema: true | |
| enable_xformers_memory_efficient_attention: false | |
| allow_tf32: true | |
| adam_beta1: 0.9 | |
| adam_beta2: 0.999 | |
| adam_weight_decay: 0.01 | |
| adam_epsilon: 1.0e-08 | |
| max_grad_norm: 1.0 | |
| logging_dir: logs | |
| mixed_precision: fp16 | |
| validation_timesteps: 128 | |
| validation_fps: ${globals.target_fps} | |
| validation_frames: ${globals.target_nframes} | |
| validation_lvefs: | |
| - 0.0 | |
| - 0.4 | |
| - 0.7 | |
| - 1.0 | |
| validation_guidance: 1.0 | |
| validation_steps: 1500 | |
| report_to: wandb | |
| checkpointing_steps: 10000 | |
| checkpoints_total_limit: 100 | |
| resume_from_checkpoint: null | |
| tracker_project_name: echosyn | |
| seed: 42 | |
| num_train_epochs: 893 | |