model:
  _target_: prismwam.models.wan22.rgb_depth_action_depth_parent.create_prismwam_rgb_depth_action
  model_id: Wan-AI/Wan2.2-TI2V-5B
  tokenizer_model_id: Wan-AI/Wan2.1-T2V-1.3B
  tokenizer_max_len: 128
  load_text_encoder: false
  redirect_common_files: true
  architecture_name: PrismWAM
  rgb_input_size:
  - 384
  - 320
  rgb_target_mode: f8
  rgb_depth_attention_mode: rgb_to_depth_to_action
  rgb_fast: true
  rgb_history_num_frames: 5
  rgb_history_frame_stride: 4
  share_rgb_depth_blocks: true
  video_use_text_embedding: true
  proprio_dim: 14
  mot_checkpoint_mixed_attn: true
  skip_depth_dit_load_from_pretrain: true
  skip_rgb_dit_load_from_pretrain: true
  skip_action_dit_load_from_pretrain: true
  checkpoint_init_only: true
  depth_vae_require_latent_stats: true
  depth_dit_checkpoint_path: null
  depth_vae_checkpoint_path: depth_vae.pt
  rgb_vae_checkpoint_path: rgb_vae.safetensors
  action_dit_pretrained_path: null
  depth_dit_config:
    patch_size:
    - 1
    - 2
    - 2
    in_dim: 48
    hidden_dim: 3072
    ffn_dim: 14336
    freq_dim: 256
    text_dim: 4096
    out_dim: 48
    num_heads: 24
    attn_head_dim: 128
    num_layers: 30
    eps: 1.0e-06
    seperated_timestep: true
    require_vae_embedding: false
    fuse_vae_embedding_in_latents: true
    use_gradient_checkpointing: true
    video_attention_mask_mode: first_frame_causal
  rgb_dit_config:
    patch_size:
    - 1
    - 2
    - 2
    in_dim: 48
    hidden_dim: 3072
    ffn_dim: 14336
    freq_dim: 256
    text_dim: 4096
    out_dim: 48
    num_heads: 24
    attn_head_dim: 128
    num_layers: 30
    eps: 1.0e-06
    seperated_timestep: true
    require_vae_embedding: false
    fuse_vae_embedding_in_latents: true
    use_gradient_checkpointing: true
    video_attention_mask_mode: first_frame_causal
  action_dit_config:
    action_dim: 14
    hidden_dim: 1024
    ffn_dim: 4096
    num_heads: 24
    attn_head_dim: 128
    num_layers: 30
    text_dim: 4096
    freq_dim: 256
    eps: 1.0e-06
    use_gradient_checkpointing: true
  rgb_scheduler:
    train_shift: 5.0
    infer_shift: 5.0
    num_train_timesteps: 1000
  depth_scheduler:
    train_shift: 5.0
    infer_shift: 5.0
    num_train_timesteps: 1000
  action_scheduler:
    train_shift: 5.0
    infer_shift: 5.0
    num_train_timesteps: 1000
  loss:
    lambda_rgb: 1.0
    lambda_depth: 1.0
    lambda_action: 1.0
  depth_parent_expected_step: 15000
  action_init_source: wan
  lr_schedule_steps: 15000
data:
  train:
    video_size:
    - 384
    - 320
    num_video_frames: 9
    frame_stride: 4
    rgb_history_num_frames: 5
    rgb_history_frame_stride: 4
    cameras:
    - head_camera
    - left_camera
    - right_camera
    layout: robotwin
    depth_min: 0.0
    depth_max: 1500.0
    action_horizon: 32
    action_dim: 14
    action_start_offset: 1
    proprio_start_offset: 0
    context_len: 128
    text_encoder_id: wan22ti2v5b
    prompt_template: 'A video recorded from a robot''s point of view executing the
      following instruction: {task}'
    task_names: null
    rgb_target_mode: f8
deployment:
  action_horizon: 32
  replan_steps: 24
  num_inference_steps: 10
  inference_shifts:
    depth: 5.0
    action: 5.0
  benchmark: robotwin
  portable_bundle: true
  statistics_path: action_stats.pt