| { | |
| "run_id": "0119_QwenLatent_libero_bs64_addnorm_chunk15", | |
| "run_root_dir": "./runs", | |
| "seed": 42, | |
| "trackers": [ | |
| "jsonl", | |
| "wandb" | |
| ], | |
| "wandb_entity": "timsty", | |
| "wandb_project": "vla_jepa", | |
| "is_debug": false, | |
| "framework": { | |
| "name": "QwenLatent", | |
| "qwenvl": { | |
| "base_vlm": "/data2/timsty/checkpoint/Qwen/Qwen3-VL-2B-Instruct", | |
| "attn_implementation": "flash_attention_2", | |
| "vl_hidden_dim": 2048, | |
| "num_data_tokens": 32 | |
| }, | |
| "action_model": { | |
| "ckpt_path": "/data2/timsty/code/LearnLatent/runs/0119_ActionModel_droid_furniture_libero_rt1_bridge_addnorm/final_model/pytorch_model.pt", | |
| "action_size": 32, | |
| "state_size": 32, | |
| "hidden_size": 1024, | |
| "intermediate_size": 3072, | |
| "dataset_vocab_size": 256, | |
| "num_data_tokens": 8, | |
| "mask_ratio": 0.25, | |
| "state_drop_prob": 0.5, | |
| "min_action_len": 5, | |
| "num_encoder_layers": 28, | |
| "num_decoder_layers": 28, | |
| "num_attention_heads": 16, | |
| "num_key_value_heads": 8, | |
| "head_dim": 128, | |
| "max_position_embeddings": 2048, | |
| "max_action_chunk_size": 50, | |
| "rms_norm_eps": 1e-06, | |
| "attention_dropout": 0.0, | |
| "use_vae_reparameterization": false | |
| } | |
| }, | |
| "datasets": { | |
| "vla_data": { | |
| "dataset_py": "lerobot_datasets", | |
| "data_root_dir": "/data2/timsty/dataset/IPEC-COMMUNITY", | |
| "data_mix": "libero_all", | |
| "action_type": "delta_ee", | |
| "CoT_prompt": "Task: {instruction}. Robot State: {state_str}. What are the next {chunk_size} actions to take?", | |
| "default_image_resolution": [ | |
| 3, | |
| 224, | |
| 224 | |
| ], | |
| "per_device_batch_size": 64, | |
| "load_all_data_for_training": true, | |
| "obs": [ | |
| "image_0" | |
| ], | |
| "image_size": [ | |
| 224, | |
| 224 | |
| ], | |
| "video_backend": "torchvision_av", | |
| "load_past": false, | |
| "load_future": false, | |
| "chunk_size": 15 | |
| } | |
| }, | |
| "trainer": { | |
| "epochs": 100, | |
| "max_train_steps": 10000, | |
| "num_warmup_steps": 500, | |
| "save_interval": 1000, | |
| "eval_interval": 50, | |
| "learning_rate": { | |
| "base": 5e-05, | |
| "qwen_vl_interface": 5e-05, | |
| "action_model": 5e-05 | |
| }, | |
| "lr_scheduler_type": "cosine_with_min_lr", | |
| "scheduler_specific_kwargs": { | |
| "min_lr": 5e-07 | |
| }, | |
| "freeze_modules": "", | |
| "loss_scale": { | |
| "align_loss": 1.0, | |
| "recon_loss": 0.1 | |
| }, | |
| "warmup_ratio": 0.1, | |
| "weight_decay": 0.0, | |
| "logging_frequency": 10, | |
| "gradient_clipping": 5.0, | |
| "gradient_accumulation_steps": 1, | |
| "optimizer": { | |
| "name": "AdamW", | |
| "betas": [ | |
| 0.9, | |
| 0.95 | |
| ], | |
| "eps": 1e-08, | |
| "weight_decay": 1e-08 | |
| }, | |
| "is_resume": false, | |
| "resume_epoch": null, | |
| "resume_step": null, | |
| "enable_gradient_checkpointing": true, | |
| "enable_mixed_precision_training": true | |
| }, | |
| "output_dir": "./runs/0119_QwenLatent_libero_bs64_addnorm_chunk15" | |
| } |