Add runtime aliases and CogACT base metadata
Browse files- config.json +63 -0
config.json
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_dim": 7,
|
| 3 |
+
"action_model_type": "DiT-L",
|
| 4 |
+
"consolidate_type": "tome",
|
| 5 |
+
"data_root_dir": "/training-results/haodong/Real/memoryvla_rlds",
|
| 6 |
+
"dataloader_type": "stream",
|
| 7 |
+
"ema_decay": 0.999,
|
| 8 |
+
"fusion_type": "gate",
|
| 9 |
+
"future_action_window_size": 15,
|
| 10 |
+
"group_size": 16,
|
| 11 |
+
"hf_token": "HF_TOKEN",
|
| 12 |
+
"image_aug": false,
|
| 13 |
+
"image_aug_mode": "spatial",
|
| 14 |
+
"is_resume": false,
|
| 15 |
+
"load_all_data_for_training": true,
|
| 16 |
+
"mem_length": 256,
|
| 17 |
+
"per_token_size": 256,
|
| 18 |
+
"pretrained_checkpoint": "/training-results/haodong/MemoryVLA/code/pretrained/CogACT-Large/checkpoints/CogACT-Large.pt",
|
| 19 |
+
"repeated_diffusion_steps": 4,
|
| 20 |
+
"resume_epoch": 0,
|
| 21 |
+
"resume_step": 0,
|
| 22 |
+
"retain_optimizer_checkpoints": 2,
|
| 23 |
+
"retrieval_layers": 2,
|
| 24 |
+
"run_id": "piper-color-sorting-frozen16-ema-real21-20260823-053817",
|
| 25 |
+
"run_id_note": null,
|
| 26 |
+
"run_root_dir": "/training-results/haodong/MemoryVLA/runs/piper-color-sorting",
|
| 27 |
+
"save_interval": 2000,
|
| 28 |
+
"save_on_terminate": true,
|
| 29 |
+
"seed": 42,
|
| 30 |
+
"trackers": [
|
| 31 |
+
"jsonl",
|
| 32 |
+
"wandb"
|
| 33 |
+
],
|
| 34 |
+
"update_fused": false,
|
| 35 |
+
"use_ema": true,
|
| 36 |
+
"use_timestep_pe": true,
|
| 37 |
+
"vla": {
|
| 38 |
+
"base_vlm": "prism-dinosiglip-224px+7b",
|
| 39 |
+
"data_mix": "custom_finetuning",
|
| 40 |
+
"enable_gradient_checkpointing": true,
|
| 41 |
+
"enable_mixed_precision_training": true,
|
| 42 |
+
"epochs": 100,
|
| 43 |
+
"expected_world_size": 16,
|
| 44 |
+
"freeze_llm_backbone": true,
|
| 45 |
+
"freeze_vision_backbone": true,
|
| 46 |
+
"global_batch_size": 32,
|
| 47 |
+
"learning_rate": 2e-05,
|
| 48 |
+
"lr_scheduler_type": "linear-warmup+cosine-decay",
|
| 49 |
+
"max_grad_norm": 1.0,
|
| 50 |
+
"max_steps": 20000,
|
| 51 |
+
"per_device_batch_size": 2,
|
| 52 |
+
"reduce_in_full_precision": true,
|
| 53 |
+
"shuffle_buffer_size": 64,
|
| 54 |
+
"train_strategy": "fsdp-full-shard",
|
| 55 |
+
"type": "prism-dinosiglip-224px+oxe+diffusion",
|
| 56 |
+
"unfreeze_last_llm_layer": false,
|
| 57 |
+
"vla_id": "prism-dinosiglip-224px+oxe+diffusion",
|
| 58 |
+
"warmup_ratio": 0.03,
|
| 59 |
+
"weight_decay": 0.0
|
| 60 |
+
},
|
| 61 |
+
"wandb_entity": "spikingtransformer",
|
| 62 |
+
"wandb_project": "Memory World Model"
|
| 63 |
+
}
|