Add training config

config.json

@@ -0,0 +1,79 @@
+{
+  "vla_name": "VITRA_Paligemma",
+  "task_name": "opentouch_keypoint_from_step140000_large",
+  "model": "vitra_paligemma2",
+  "fwd_pred_next_n": 16,
+  "seed": 42,
+  "batch_size": 1,
+  "output_root": "/home/chonghej/scratch/chonghej/VLA-HAND/runs/opentouch_step140000_adapt_touch_editor/adapt_checkpoints",
+  "log_root": "/home/chonghej/scratch/chonghej/VLA-HAND/runs/opentouch_step140000_adapt_touch_editor/adapt_logs",
+  "cache_root": "runs/opentouch_keypoint_from_step140000_adapt/cache/vitra_paligemma2",
+  "model_load_path": null,
+  "pretrain_path": "runs/gigahands_real_all_cam0_keypoints_mano_vitra3b_linked_train/checkpoints/gigahands_real_all_cam0_keypoints_mano_vitra3b_linked_stage1_TB2_B2_bf16True/checkpoints/epoch=0-step=140000.ckpt",
+  "resume": false,
+  "wandb_project": "vitra-opentouch-keypoint-adapt",
+  "wandb_entity": "",
+  "save_steps": 5000,
+  "epoch_save_interval": 100000,
+  "total_batch_size": 4,
+  "use_bf16": true,
+  "use_fov": true,
+  "untied_cognition_token": true,
+  "use_state": "DiT",
+  "loss_type": "human",
+  "train_setup": {
+    "freeze_option": "freeze_vision_encoder"
+  },
+  "state_encoder": {
+    "state_dim": 212
+  },
+  "action_model": {
+    "model_type": "DiT-B",
+    "token_size": 2304,
+    "action_dim": 192,
+    "hidden_size": 1024
+  },
+  "vlm": {
+    "type": "PaliGemmaForConditionalGeneration",
+    "name": "paligemma",
+    "pretrained_model_name_or_path": "google/paligemma2-3b-mix-224"
+  },
+  "trainer": {
+    "sharding_strategy": "shard-grad-op",
+    "strategy": "fsdp_paligemma_with_checkpointing",
+    "lr_scheduler_type": "backbone-freeze-warmup",
+    "gradient_clip_val": 1.0,
+    "learning_rate": 1e-05,
+    "weight_decay": 0.1,
+    "max_epochs": 100000,
+    "max_steps": 10000,
+    "reduce_in_full_precision": true,
+    "enable_mixed_precision_training": false,
+    "enable_gradient_checkpointing": true,
+    "action_model_learning_rate": 0.0001,
+    "llm_freeze_step": 5000,
+    "warmup_ratio": null
+  },
+  "train_dataset": {
+    "data_root_dir": "datasets/vitra_opentouch_keypoint_full_text_aligned",
+    "augmentation": false,
+    "set_none_ratio": 0.0,
+    "data_mix": "opentouch_keypoint_train",
+    "num_workers": 2,
+    "prefetch_factor": null,
+    "flip_augmentation": 1.0,
+    "action_type": "keypoints",
+    "use_rel": false,
+    "clip_len": null,
+    "normalization": true,
+    "state_mask_prob": 0.0,
+    "statistics_dataset_name": "opentouch_keypoint_train"
+  },
+  "repeated_diffusion_steps": 8,
+  "config": "/home/chonghej/scratch/chonghej/VLA-HAND/vitra/configs/human_pretrain_opentouch_keypoint_from_step140000.json",
+  "data_mix": null,
+  "debug": false,
+  "num_workers": 2,
+  "prefetch_factor": null,
+  "no_save_checkpoint": false
+}