training_config.yaml

seed: 1234
model:
  type: improved_ais_bilstm
  pose_dim: 596
  input_dim: 597
  temporal_dim: 4
  hidden_size: 512
  num_layers: 2
  dropout: 0.3003
  synthesis_hidden: 1024
  beta_groups: 64
data:
  clip_length: 224
  stride: 56
  val_stride: 224
  train_controller_h5: artifacts/hf_dataset/in_house_dataset/vectorized_controller_values/ef25a8e5ecd2fa86420741e5646cd785aa025c44c06f4118aab77f558c8f6981/train.h5
  train_block_keyframes_h5: artifacts/hf_dataset/in_house_dataset/vectorized_block_keyframes/train.h5
  val_controller_h5: artifacts/hf_dataset/in_house_dataset/vectorized_controller_values/ef25a8e5ecd2fa86420741e5646cd785aa025c44c06f4118aab77f558c8f6981/test.h5
  val_block_keyframes_h5: artifacts/hf_dataset/in_house_dataset/vectorized_block_keyframes/test.h5
training:
  output_dir: runs/ais_major_scratch_10k
  device: auto
  batch_size: 8
  eval_batch_size: 8
  num_workers: 0
  max_steps: 10000
  gradient_accumulation_steps: 8
  progress: true
  progress_bar: true
  progress_every_steps: 25
  dashboard: true
  train_log_every_steps: 25
  val_every_steps: 1000
  save_every_steps: 1000
  grad_clip_norm: 1.0
  lr: 0.0001
  weight_decay: 0.001
  hard_copy_keyposes_in_forward: false
  eval_only: false
  checkpoint:
    best_metric: val_loss
    save_periodic: true
    keep_every_n: 5
loss:
  velocity_weight: 0.05
  acceleration_weight: 0.01
  spectral_weight: 0.001
  gate_tv_weight: 0.001
  keypose_weight: 10.0
eval:
  write_report: false
  baselines:
  - linear
  - previous
  max_report_clips: 20
  max_report_plots: 12
  sets:
  - name: prod_test
    controller_h5: artifacts/hf_dataset/prod_test_dataset/vectorized_controller_values/ef25a8e5ecd2fa86420741e5646cd785aa025c44c06f4118aab77f558c8f6981/test.h5
    block_keyframes_h5: artifacts/hf_dataset/prod_test_dataset/vectorized_block_keyframes/test.h5
    stride: 224
    max_clips: 20