| { |
| "run_id": "robotwin_clean_implicit_jepa_b200x8_full_20260708_212013", |
| "run_root_dir": "results", |
| "seed": 42, |
| "wandb_entity": "livion", |
| "wandb_project": "m2w-vla", |
| "is_debug": false, |
| "framework": { |
| "name": "QwenSubtaskM2W_robotwin_phs", |
| "qwenvl": { |
| "base_vlm": "playground/Pretrained_models/Qwen3-VL-4B-Instruct", |
| "attn_implementation": "sdpa", |
| "vl_hidden_dim": 2048 |
| }, |
| "reasoning": { |
| "cot_reasoning_mode": "implicit_query", |
| "latent_cot_num_queries": 16, |
| "qwen_action_context_mode": "exclude_cot_prompt", |
| "lambda_cot": 0.1, |
| "generate_cot_at_inference": false, |
| "inference_cot_max_new_tokens": 96, |
| "inference_cot_use_cache": true, |
| "inference_empty_cache": false |
| }, |
| "vjepa2": { |
| "hub_repo": "facebookresearch/vjepa2:main", |
| "hub_model": "vjepa2_1_vit_large_384", |
| "hub_source": "github", |
| "num_frames": 8, |
| "image_size": 224 |
| }, |
| "jepa_predictor": { |
| "prediction_view": "wrist", |
| "num_wrist_views": 2, |
| "dropout": 0.0, |
| "use_jepa_loss": true, |
| "lambda_jepa": 0.2, |
| "jepa_predictor_depth": 4, |
| "jepa_predictor_num_heads": 8, |
| "jepa_predictor_bf16_autocast": true, |
| "jepa_predictor_activation_checkpointing": true, |
| "view_embed_init_std": 0.02 |
| }, |
| "policy_context": { |
| "use_state": false, |
| "wrist_action_context_tokens": 32, |
| "wrist_action_context_heads": 8, |
| "wrist_action_context_layers": 1, |
| "detach_wrist_action_context": true |
| }, |
| "action_model": { |
| "action_model_type": "DiT-B", |
| "gradient_checkpointing": true, |
| "action_hidden_dim": 1024, |
| "hidden_size": 1024, |
| "add_pos_embed": true, |
| "max_seq_len": 1024, |
| "action_dim": 14, |
| "state_dim": 14, |
| "action_horizon": 16, |
| "future_action_window_size": 15, |
| "past_action_window_size": 0, |
| "noise_beta_alpha": 1.5, |
| "noise_beta_beta": 1.0, |
| "noise_s": 0.999, |
| "num_timestep_buckets": 1000, |
| "num_inference_timesteps": 4, |
| "num_target_vision_tokens": 32, |
| "diffusion_model_cfg": { |
| "cross_attention_dim": 2560, |
| "dropout": 0.2, |
| "final_dropout": true, |
| "interleave_self_attention": true, |
| "norm_type": "ada_norm", |
| "num_layers": 16, |
| "output_dim": 1024, |
| "positional_embeddings": null |
| } |
| } |
| }, |
| "datasets": { |
| "vla_data": { |
| "dataset_py": "subtask_m2w_robotwin_phs_datasets", |
| "data_root_dir": "playground/Datasets/RoboTwin", |
| "data_mix": "robotwin_clean_h16", |
| "action_type": "abs_qpos", |
| "action_mode": "abs", |
| "per_device_batch_size": 16, |
| "load_all_data_for_training": true, |
| "video_backend": "torchvision_av", |
| "num_workers": 3, |
| "prefetch_factor": 2, |
| "include_state": false, |
| "image_size": [ |
| 224, |
| 224 |
| ], |
| "obs_image_size": [ |
| 224, |
| 224 |
| ], |
| "jepa_prediction_view": "wrist", |
| "num_wrist_views": 2, |
| "vlm_include_wrist_view": true, |
| "wrist_history_frames": 8, |
| "load_wrist_future_views": true, |
| "future_wrist_k": 16, |
| "future_wrist_small_gap": 2, |
| "subtask_label_dir": "playground/Datasets/subtask_m2w_labels_robotwin_qwen3vl32b_full_2500_repair_candidate_v13_style_unify_20260707", |
| "cot_prompt_version": "robotwin_tri_v1", |
| "CoT_prompt": "Your task is: {instruction} Observation: using the current high camera, left wrist camera, and right wrist camera images, identify the robot's current manipulation step.\nOutput exactly:\nSubtask: ...\nReasoning: ...\nWrist: left=...; right=...\nDo not add extra text.", |
| "qwen_image_augmentation": true, |
| "qwen_image_augmentation_prob": 0.25, |
| "qwen_image_augmentation_strength": 0.5 |
| } |
| }, |
| "trainer": { |
| "max_train_steps": 150000, |
| "num_warmup_steps": 5000, |
| "is_resume": false, |
| "pretrained_checkpoint": null, |
| "reload_modules": null, |
| "resume_optimizer_state": true, |
| "train_qwen_vl": true, |
| "find_unused_parameters": false, |
| "save_interval": 10000, |
| "eval_interval": 1000, |
| "logging_frequency": 100, |
| "repeated_diffusion_steps": 8, |
| "learning_rate": { |
| "base": 2.5e-05, |
| "jepa_predictor": 5e-05, |
| "wrist_context_adapter": 5e-05, |
| "action_model": 0.0001, |
| "vlm": 1e-05 |
| }, |
| "lr_scheduler_type": "cosine_with_min_lr", |
| "scheduler_specific_kwargs": { |
| "min_lr": 1e-06 |
| }, |
| "gradient_clipping": 1.0, |
| "gradient_accumulation_steps": 1, |
| "optimizer": { |
| "name": "AdamW", |
| "betas": [ |
| 0.9, |
| 0.95 |
| ], |
| "eps": 1e-08, |
| "weight_decay": 1e-08 |
| } |
| }, |
| "output_dir": "results/robotwin_clean_implicit_jepa_b200x8_full_20260708_212013" |
| } |
|
|
