configs.json

{
  "evaluate_during_training": true,
  "record_evaluation_video": false,
  "compile_model": false,
  "bytedance_ckpt_path": "./snapshot_ABC.pt",
  "load_bytedance_ckpt": false,
  "load_epoch": 0,
  "save_epochs": 1,
  "num_epochs": 20,
  "num_sequences": 1000,
  "ep_len": 360,
  "print_steps": 100,
  "lr_max": 0.0009,
  "weight_decay": 0.0001,
  "arm_loss_ratio": 100,
  "num_warmup_epochs": 1,
  "bs_per_gpu": 19,
  "workers_per_gpu": 7,
  "prefetch_factor": 2,
  "gradient_accumulation_steps": 3,
  "rgb_static_pad": 10,
  "rgb_gripper_pad": 4,
  "rgb_shape": 224,
  "patch_size": 16,
  "rgb_mean": [0.485, 0.456, 0.406],
  "rgb_std": [0.229, 0.224, 0.225],
  "LMDB_path": "../calvin_lmdb/",
  "save_path": "./Save/",
  "mae_ckpt": "./mae_pretrain_vit_base.pth",
  "embed_dim": 384,
  "n_layer": 12,
  "n_head": 12,
  "activation_function": "relu",
  "dropout": 0.1,
  "n_positions": 1024,
  "device": "cuda",
  "resampler_depth": 3,
  "resampler_dim_head": 128,
  "resampler_heads": 4,
  "resampler_num_media_embeds": 1,
  "resampler_num_latents": 9,
  "seq_len": 10,
  "chunk_size": 10,
  "test_chunk_size": 1,
  "skip_frame": 3,
  "action_mode": "ee_rel_pose",
  "act_dim": 7,
  "state_dim": 7,
  "use_hand_rgb": true,
  "clip_backbone": "ViT-B/32",
  "act_pred": true,
  "fwd_pred": true,
  "fwd_pred_hand": true,
  "fwd_pred_next_n": 3,
  "without_norm_pixel_loss": false,
  "img_feat_dim": 768,
  "patch_feat_dim": 768,
  "lang_feat_dim": 512
}