run_id: 1007_Pi05_stack_cube_red_yellow_subtask
run_root_dir: ./results/Checkpoints/vla
seed: 42
trackers:
- jsonl
is_debug: false
framework:
  name: Pi05
  model:
    model_path: ${oc.env:PI05_MODEL_PATH,playground/Pretrained_models/pi05_base_pytorch}
    tokenizer_path: ${oc.env:PI05_TOKENIZER_PATH,/inspire/ssd/project/robot3d/mazipei-253107140027/.cache/openpi/big_vision/paligemma_tokenizer.model}
    dtype: bfloat16
    paligemma_variant: gemma_2b
    action_expert_variant: gemma_300m
    vision_variant: So400m/14
    max_token_len: 200
    num_images: 3
    fast_tokenizer_path: playground/Pretrained_models/fast
    fast_max_tokens: null
    action_ce_weight: 0.01
    subtask_enabled: true
    subtask_ce_weight: 0.01
    subtask_max_tokens: 64
    infer_mode: flow_with_subtask
  action_model:
    action_dim: 14
    state_dim: 14
    model_action_dim: 32
    action_horizon: 50
    num_inference_steps: 10
    rtc_training_max_delay: 0
datasets:
  vla_data:
    normalization_mode: min_max
    dataset_py: vla_datasets
    subtask_key: subtask_generation
    disable_state: true
    data_root_dir: playground/Datasets/LEROBOT_LIBERO_DATA
    data_mix: stack_cube_red_yellow
    per_device_batch_size: 8
    split_strategy: episode_ratio
    eval_ratio: 0.02
    split_seed: 42
    eval_per_device_batch_size: 4
    num_workers: 8
    prefetch_factor: 4
    load_all_data_for_training: true
    action_horizon: ${framework.action_model.action_horizon}
    image_size:
    - 224
    - 224
    state_pad_size: ${framework.action_model.state_dim}
    action_pad_size: ${framework.action_model.action_dim}
    use_future_frames: false
    future_frame_stride: 1
trainer:
  global_batch_size: 64
  max_train_steps: 30000
  num_warmup_steps: 1000
  save_interval: 2000
  eval_interval: 2000
  learning_rate:
    base: 2.5e-05
  lr_scheduler_type: cosine_with_min_lr
  scheduler_specific_kwargs:
    min_lr: 0.0
  freeze_modules: null
  loss_scale:
    vla: 1.0
  max_grad_norm: 1.0
  weight_decay: 0.0
  logging_frequency: 10
  gradient_clipping: 1.0
  gradient_accumulation_steps: 1
  optimizer:
    name: AdamW
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-10
  enable_gradient_checkpointing: true
  enable_compile: true
  resume_from_checkpoint: latest
  save_format: pt
output_dir: ./results/Checkpoints/vla/1007_Pi05_stack_cube_red_yellow_subtask