run_id: Wan2.2-5B-PI-SIMPLE
run_root_dir: ./playground/Checkpoints
seed: 42
wandb_entity: null
wandb_project: starVLA_SIMPLE
is_debug: false
version_id: '0.21'
framework:
  name: StarVLA-WM-Wan-PI
  world_model:
    base_wm: Wan-AI/Wan2.2-TI2V-5B-Diffusers
    extract_layers:
    - -1
  qwenvl:
    base_vlm: Wan-AI/Wan2.2-TI2V-5B-Diffusers
    vl_hidden_dim: 1024
    num_vl_layers: 30
  action_model:
    action_model_type: LayerwiseFM
    action_dim: 36
    state_dim: 32
    future_action_window_size: 29
    past_action_window_size: 0
    repeated_diffusion_steps: 2
    num_inference_timesteps: 4
    add_pos_embed: true
    max_seq_len: 1024
    num_target_vision_tokens: 32
    noise_beta_alpha: 1.5
    noise_beta_beta: 1.0
    noise_s: 0.999
    num_timestep_buckets: 1000
    diffusion_model_cfg:
      cross_attention_dim: 1024
      action_dit_hidden_dim: 1024
      dropout: 0.2
      final_dropout: true
      interleave_self_attention: true
      norm_type: ada_norm
      positional_embeddings: null
      attention_head_dim: 64
      num_layers: 30
      input_embedding_dim: 2048
      num_attention_heads: 32
    action_horizon: 30
datasets:
  vlm_data:
    dataset_py: vlm_datasets
    dataformat: llava_json
    dataset_use: sharegpt4v_coco
    eval_dataset: sharegpt4v_coco
    data_flatten: false
    base_interval: 2
    max_pixels: 50176
    min_pixels: 784
    model_max_length: 2048
    model_type: qwen2.5vl
    per_device_batch_size: 1
  vla_data:
    dataset_py: lerobot_datasets
    include_state: true
    data_root_dir: ./playground/Datasets
    data_mix: simple_psi0_all_twelve_tasks
    action_type: abs_qpos
    action_mode: abs
    per_device_batch_size: 4
    load_all_data_for_training: true
    obs_image_size:
    - 832
    - 480
    video_backend: torchvision_av
trainer:
  max_train_steps: 120000
  num_warmup_steps: 5000
  save_interval: 10000
  eval_interval: 1000
  learning_rate:
    base: 1.0e-05
    wm_projector: 0.0001
    action_model: 0.0001
  lr_scheduler_type: cosine_with_min_lr
  scheduler_specific_kwargs:
    min_lr: 5.0e-07
  freeze_modules: backbone.text_encoder,backbone.vae
  loss_scale:
    vla: 1.0
    vlm: 0.0
  max_grad_norm: 1.0
  weight_decay: 0.0
  logging_frequency: 10
  gradient_clipping: 1.0
  gradient_accumulation_steps: 1
  gradient_checkpointing: true
  optimizer:
    name: AdamW
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  repeated_diffusion_steps: 2
  is_resume: true
config_yaml: ./examples/simBenchmarks/SIMPLE/train_files/starvla_simple_psi0_wanpi.yaml
output_dir: ./playground/Checkpoints/Wan2.2-5B-PI-SIMPLE