stage_1:
  ckpt_path: /iopsstor/scratch/cscs/hongsu/seonghu/checkpoints/track4world_da3.pth
  encoder_input_size: 224
  normalization_stat_path: null
da3_finetune:
  enabled: true
  freeze_blocks_before: 13
  n_action_steps: 2
  n_views: 2
  use_temporal_embed: false
  action_input_rate: 0.4
action_head:
  type: mlp_resnet
  input_dim: 1536
  hidden_dim: 1536
  n_dims: 7
  chunk_size: 8
  num_blocks: 2
  pool_mode: mean
regularization:
  lambda_feat: 0.0
  layer_weight_min: 0.5
  adaptive_lambda: false
  lambda_depth: 3.0
  depth_grad_weight: 1.0
  depth_decode_chunk_size: 8
  lambda_camera: 0.0
  teacher_depth_fallback: true
  skip_depth_if_no_gt: false
  lambda_path_b_deep_feat_reg: 0.0
  deep_feat_reg_layer_weight_min: 0.5
proprioception:
  enabled: true
  proprio_dim: 7
  hidden_dim: 256
predictor:
  enabled: true
  type: shallow12_ar
  d_model: 1024
  depth: 12
  num_heads: 16
  ffn_ratio: 4.0
  dropout: 0.0
  num_patches_per_view: 256
  use_language: true
  language_encoder_type: t5
  language_dim: 768
  language_len: 77
  clip_model: openai/clip-vit-large-patch14
  t5_model: google-t5/t5-base
  condition_mode: concat
  input_proj_norm: ln
  cache_token_embeddings: true
  language_cache_max_entries: 4096
  language_cache_device: cpu
  lambda_feat_future: 1.0
  lambda_feat_current: 0.0
  lambda_proprio_future: 0.0
  lambda_sigreg: 0.0
  feature_loss_type: l1
  feature_target_mode: future
  H_choices:
  - 1
  H_weights:
  - 1.0
  num_register_tokens: 0
  deep_gradient_checkpointing: false
  use_proprio_head: false
training:
  global_batch_size: 512
  micro_batch_size: 32
  grad_accum_steps: 1
  min_train_batches_per_epoch: 128
  base_lr: 5.16e-05
  head_lr_mult: 10.0
  predictor_lr_mult: 10.0
  adam_eps: 1.0e-06
  adam_beta1: 0.9
  adam_beta2: 0.95
  weight_decay: 0.0
  lambda_action: 9.0
  lambda_action_direct: 0.0
  lambda_action_refine: 1.0
  epochs: 500000
  clip_grad: 1.0
  warmup_steps: 500
  max_steps: 100000
  min_lr_ratio: 1.0
  log_every: 25
  ckpt_every: 2000
  vis_every: 200
  visualization:
    log_unified_depth: true
  eval_every: 2000
  global_seed: 42
  num_workers: 8
  prefetch_factor: 4
  persistent_workers: true
  bf16: true
  compile: false
  distributed_timeout_minutes: 120
  preemption_check_every_steps: 0
dataset:
  type: openx
  openx_root: /iopsstor/scratch/cscs/hongsu/seonghu/data/lerobot_franka
  datasets:
  - name: franka_pickplace
  - name: franka_chocomilk
  - name: franka_kitchen
  - name: franka_pot
  sampling_strategy: dataset_weighted
  stats_dir: /iopsstor/scratch/cscs/hongsu/seonghu/checkpoints/realrobot/franka_multitask_openvla_noidle_stats_0034000/stats
  image_size:
  - 224
  - 224
  future_steps: 1
  chunk_size: 8
  include_current_action: true
  uniform_action_sampling: true
  proprio_dim: 7
  n_views: 2
  source_fps: 15
  action_frame: base_delta
  download_videos: false
  lazy_download: false
  force_cache_sync: false
  enable_arrow_cache: false
  da3_input_rotate180: false
  da3_input_hflip: false
  da3_input_vflip: false
  openpi_libero_augment: true
  openpi_base_crop_scale: 0.9486832980505138
  openpi_base_rotate_degrees: 5.0
  color_jitter_brightness: 0.3
  color_jitter_contrast: 0.4
  color_jitter_saturation: 0.5
  color_jitter_hue: 0.05
  image_augmentation:
    profile: cosmos_policy_strong
    train:
      random_resized_crop_area: 0.9
      base_only_rotation_degrees: 5.0
      color_jitter:
        brightness: 0.3
        contrast: 0.4
        saturation: 0.5
        hue: 0.05
      jpeg:
        enabled: true
        quality: 95
    eval:
      center_crop_area: 0.9
      jpeg:
        enabled: true
        quality: 95
  eval_ratio: 0.05
  action_stats_samples: 0
  proprio_stats_samples: 0
wandb:
  project: GAM_realrobot
use_dit: false