actor:
  model:
    model_type: openpi_pytorch
    precision: fp32
    is_lora: false
    lora_rank: 32
    num_action_chunks: 20
    action_dim: 10
    use_proprio: true
    num_steps: 5
    add_value_head: false
    openpi:
      task: sft
      config_name: pi05_singlefranka_tcp_rot6d
      num_images_in_input: 3
      action_chunk: 20
      action_env_dim: 10
      model_action_dim: 32
      paligemma_variant: gemma_2b
      action_expert_variant: gemma_300m
      max_token_len: 200
      discrete_state_input: false
      assets_dir: /mnt/workspace/embody-h2/users/zhaozihao/DAgger/assets/pi05_singlefranka_tcp_rot6d/umi_left_franka_tcp_rot6d
      asset_id: fold_towel_umi-50plusEgodagger-franka_post_train_80_20
    model_path: /mnt/workspace/embody-h2/users/zhaozihao/checkpoint_archive/fold_towel_umi50plusEgodagger_sft4000_weights_only
  fsdp_config:
    strategy: fsdp
    sharding_strategy: full_shard
    gradient_checkpointing: true
    cpu_offload: false
    offload_pin_memory: false
    reshard_after_forward: true
    enable_gradient_accumulation: true
    forward_prefetch: false
    limit_all_gathers: false
    backward_prefetch: null
    use_orig_params: true
    use_liger_kernel: false
    mixed_precision:
      param_dtype: bf16
      reduce_dtype: fp32
      buffer_dtype: fp32
    amp_autocast:
      enabled: false
      precision: bf16
    grad_scaler:
      enabled: false
      init_scale: null
      growth_interval: null
    gradient_checkpointing_use_reentrant: false
  group_name: ActorGroup
  training_backend: fsdp
  micro_batch_size: 16
  eval_batch_size: 1
  global_batch_size: 128
  seed: 0
  optim:
    lr: 2.5e-06
    adam_beta1: 0.9
    adam_beta2: 0.95
    adam_eps: 1.0e-08
    weight_decay: 1.0e-10
    clip_grad: 1.0
    total_training_steps: 20000
    lr_warmup_steps: 500
    lr_scheduler: openpi_cosine
    num_cycles: 0.5
    min_lr: 2.5e-07
cluster:
  num_nodes: 1
  component_placement:
    actor: 0-3
runner:
  task_type: sft
  logger:
    log_path: /mnt/workspace/embody-h2/users/zhaozihao/DAgger/logs/fold_towel_umi-50plusEgodagger-franka-1
    project_name: rlinf
    experiment_name: umi_left_franka_tcp_rot6d_openpi_pytorch_post-train-weights-only8-2
    logger_backends:
    - tensorboard
  max_epochs: -1
  max_steps: 20000
  val_check_interval: -1
  save_interval: 2000
  resume_dir: null
  checkpoint_offload_dir: /mnt/workspace/embody-h2/users/zhaozihao/checkpoint_archive
  per_worker_log: false
  per_worker_log_path: null
data:
  train_data_paths: dataset/fold_towel_umi-50plusEgodagger-franka
  post_train: true
  post_train_data_path: /mnt/workspace/embody-h2/users/zhaozihao/DAgger/egodagger_dataset/fold_towel_umi-egodagger2-franka
  post_train_demo_ratio: 0.8
  post_train_correction_ratio: 0.2
  repo_id: ''
  num_workers: 4
algorithm:
  adv_type: gae
reward:
  use_reward_model: false
critic:
  use_critic_model: false