dirs:
  deps: ./deps
  raw_data: ./data
  outputs: ./outputs
  checkpoints: ./checkpoints
exp_name: ldf_263
seed: 1234
debug: false
train: true
save_dir: ${dirs.outputs}
resume_ckpt: null
test_ckpt: ${dirs.checkpoints}/ldf_263/model.ckpt
test_vae_ckpt: ${dirs.checkpoints}/vae_263/model.ckpt
test_vae:
  target: models.vae_wan.VAEWanModel
  ema_decay: 0.99
  params:
    input_dim: 263
    z_dim: 4
test_setting:
  render: true
  recover_dim: 263
  BABEL:
    compare_folders:
    - ${dirs.raw_data}/BABEL/HumanML3D263/animations
    compare_names:
    - Ground Truth
  HumanML3D:
    compare_folders:
    - ${dirs.raw_data}/HumanML3D/HumanML3D263/animations
    compare_names:
    - Ground Truth
val_repeat: 1
logger:
  wandb:
    wandb_key: ${oc.env:WANDB_API_KEY,null}
    project: FloodDiffusion2
    entity: ${oc.env:WANDB_ENTITY,null}
trainer:
  max_steps: 300000
  accelerator: gpu
  devices: 1
  log_every_n_steps: 100
  precision: bf16-mixed
  num_nodes: 1
validation:
  validation_steps: 5000
  test_steps: 5000
  save_every_n_steps: 5000
  save_top_k: 100
metrics:
  dim: 263
  t2m:
    target: metrics.HumanML3D263.t2m.T2MMetrics
    fid_target: original
    params:
      evaluate_text: true
      metric_mean_path: ${dirs.deps}/t2m/meta/mean.npy
      metric_std_path: ${dirs.deps}/t2m/meta/std.npy
      wordvectorizer:
        target: metrics.HumanML3D263.word_vectorizer.WordVectorizer
        params:
          meta_root: ${dirs.deps}/glove
          prefix: our_vab
        max_text_len: 20
      textencoder:
        target: metrics.HumanML3D263.t2m_evaluator.TextEncoderBiGRUCo
        ckpt: ${dirs.deps}/t2m/humanml3d/text_encoder.pt
        params:
          word_size: 300
          pos_size: 15
          hidden_size: 512
          output_size: 512
      moveencoder:
        target: metrics.HumanML3D263.t2m_evaluator.MovementConvEncoder
        ckpt: ${dirs.deps}/t2m/humanml3d/movement_encoder.pt
        params:
          input_size: 259
          hidden_size: 512
          output_size: 512
      motionencoder:
        target: metrics.HumanML3D263.t2m_evaluator.MotionEncoderBiGRUCo
        ckpt: ${dirs.deps}/t2m/humanml3d/motion_encoder.pt
        params:
          input_size: 512
          hidden_size: 1024
          output_size: 512
data:
  target: datasets.multi.MultiDataset
  collate_fn: datasets.multi.collate_fn
  train_bs: 32
  val_bs: 16
  test_bs: 16
  num_workers: 8
  datasets:
  - target: datasets.babel.BabelDataset
    train_meta_paths:
    - path: ${dirs.raw_data}/BABEL/HumanML3D263/train.txt
      name: BABEL
    val_meta_paths: []
    test_meta_paths:
    - path: ${dirs.raw_data}/BABEL/HumanML3D263/test_min.txt
      name: BABEL
    feature_path: new_joint_vecs
    token_path: TOKENS_20260113_233726_vae_wan_z4_2250000
    text_path: texts
    random_length: 0
    min_length: 5
    max_length: 99999
    window_length: 200
    feature_fps: 20
    token_fps: 5
  - target: datasets.humanml3d.HumanML3DDataset
    train_meta_paths:
    - path: ${dirs.raw_data}/HumanML3D/HumanML3D263/train.txt
      name: HumanML3D
    val_meta_paths:
    - path: ${dirs.raw_data}/HumanML3D/HumanML3D263/test.txt
      name: HumanML3D
    test_meta_paths:
    - path: ${dirs.raw_data}/HumanML3D/HumanML3D263/test_min.txt
      name: HumanML3D
    feature_path: new_joint_vecs
    token_path: TOKENS_20260113_233726_vae_wan_z4_2250000
    text_path: texts
    random_length: 0
    min_length: 40
    max_length: 200
    stream_mode: true
model:
  target: models.diffusion_forcing_wan.DiffForcingWanModel
  ema_decay: 0.99
  params:
    schedule_config:
      noise_type: linear
      chunk_size: 5
      steps: 10
      sigma_type: zero
      sigma_scale: 1.0
      random_epsilon: 0.05
      train_n_windows: 4
    text_config:
      len: 512
      dim: 4096
      checkpoint_path: ${dirs.deps}/t5_umt5-xxl-enc-bf16/models_t5_umt5-xxl-enc-bf16.pth
      tokenizer_path: ${dirs.deps}/t5_umt5-xxl-enc-bf16/google/umt5-xxl
    input_dim: 4
    mean_path: ${dirs.checkpoints}/ldf_263/assets/Mean.npy
    std_path: ${dirs.checkpoints}/ldf_263/assets/Std.npy
    input_keys:
      feature: token
      feature_length: token_length
      text: text
      text_end: token_text_end
    cfg_config:
      text_scale: 6.0
      null_scale: -5.0
    prediction_type: vel
    attn_type: partial
optimizer:
  target: AdamW
  params:
    lr: 0.0002
    betas:
    - 0.9
    - 0.99
    weight_decay: 0.0
    eps: 1.0e-08
lr_scheduler:
  target: CosineAnnealingLR
  params:
    T_max: 1000
    eta_min: 1.0e-06
representation: humanml3d263