# @package _global_
defaults:
  - train
  - override /data: libero
  - override /model: w0
  - _self_

posttrain_required: true
posttrain_stats: ${oc.env:W0_LIBERO_STATS,assets/libero/normalization.json}
init_checkpoint: ${oc.env:W0_LIBERO_CHECKPOINT,checkpoints/libero.pt}
resume: null
seed: 42
model:
  proprio_dim: 37
  action_horizon: 64
  action_chunk_size: 16
  max_action_offset: 15
  num_history_frames: 0
  video_plan_horizon: 96
  video_action_stride: 8
  action_dit_config:
    action_dim: 37
    num_soft_prompts: 8
    soft_prompt_domains: [libero/franka]
    domain_configs:
      - name: libero/franka
        dataset: libero
        embodiment: franka
        action_dim: 37
        proprio_dim: 37
        action_valid_indices: [7, 8, 9, 10, 11, 12, 13, 14, 15, 16]
        proprio_valid_indices: [7, 8, 9, 10, 11, 12, 13, 14, 15, 16]
        raw_eef_names: [eef_x, eef_y, eef_z, rot6d_col0_x, rot6d_col0_y, rot6d_col0_z, rot6d_col1_x, rot6d_col1_y, rot6d_col1_z, gripper_open_scale]
        action_semantics: native_delta_eef10_row_aligned

batch_size: 48
gradient_accumulation_steps: 1
learning_rate: 5.0e-5
soft_prompt_lr_scale: 1.0
num_epochs: 10
num_workers: 4
history_aware_batching: false
optimizer_group_max_numel: 200000000
save_every: 100
eval_every: 50
log_every: 10
keep_training_states: 1
checkpoint_min_free_gib: 128
freeze_video_dit: false
output_dir: ${oc.env:W0_OUTPUT_DIR,outputs}/libero/${now:%Y-%m-%d}_${now:%H-%M-%S}
wandb:
  name: internw0_libero_goal_long_random_frame