Publish WorldToken checkpoints: release files
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- 04_robocasa_scaling/README.md +26 -0
- 04_robocasa_scaling/bc_transformer_d300_seed123/config.json +462 -0
- 04_robocasa_scaling/bc_transformer_d300_seed123/run_info.json +11 -0
- 04_robocasa_scaling/scaling_n1_d1000_seed0/checkpoint_step_00100000.pt +3 -0
- 04_robocasa_scaling/scaling_n1_d1000_seed0/config.json +329 -0
- 04_robocasa_scaling/scaling_n1_d1000_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n1_d100_seed0/checkpoint_step_00010000.pt +3 -0
- 04_robocasa_scaling/scaling_n1_d100_seed0/config.json +329 -0
- 04_robocasa_scaling/scaling_n1_d100_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n1_d2900_seed0/checkpoint_step_00280000.pt +3 -0
- 04_robocasa_scaling/scaling_n1_d2900_seed0/config.json +329 -0
- 04_robocasa_scaling/scaling_n1_d2900_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n1_d300_seed0/checkpoint_step_00030000.pt +3 -0
- 04_robocasa_scaling/scaling_n1_d300_seed0/config.json +331 -0
- 04_robocasa_scaling/scaling_n1_d300_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n1_d50_seed0/checkpoint_step_00005000.pt +3 -0
- 04_robocasa_scaling/scaling_n1_d50_seed0/config.json +329 -0
- 04_robocasa_scaling/scaling_n1_d50_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n2_d1000_seed0/checkpoint_step_00100000.pt +3 -0
- 04_robocasa_scaling/scaling_n2_d1000_seed0/config.json +331 -0
- 04_robocasa_scaling/scaling_n2_d1000_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n2_d100_seed0/checkpoint_step_00010000.pt +3 -0
- 04_robocasa_scaling/scaling_n2_d100_seed0/config.json +331 -0
- 04_robocasa_scaling/scaling_n2_d100_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n2_d2900_seed0/checkpoint_step_00280000.pt +3 -0
- 04_robocasa_scaling/scaling_n2_d2900_seed0/config.json +331 -0
- 04_robocasa_scaling/scaling_n2_d2900_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n2_d300_seed0/checkpoint_step_00030000.pt +3 -0
- 04_robocasa_scaling/scaling_n2_d300_seed0/config.json +331 -0
- 04_robocasa_scaling/scaling_n2_d300_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n2_d50_seed0/checkpoint_step_00005000.pt +3 -0
- 04_robocasa_scaling/scaling_n2_d50_seed0/config.json +331 -0
- 04_robocasa_scaling/scaling_n2_d50_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n3_d1000_seed0/config.json +327 -0
- 04_robocasa_scaling/scaling_n3_d1000_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n3_d100_seed0/config.json +327 -0
- 04_robocasa_scaling/scaling_n3_d100_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n3_d2900_seed0/config.json +327 -0
- 04_robocasa_scaling/scaling_n3_d2900_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n3_d300_seed0/config.json +328 -0
- 04_robocasa_scaling/scaling_n3_d300_seed0/run_info.json +13 -0
- 04_robocasa_scaling/scaling_n3_d50_seed0/config.json +327 -0
- 04_robocasa_scaling/scaling_n3_d50_seed0/run_info.json +13 -0
- 05_token_interface/README.md +20 -0
- 05_token_interface/multitoken_n2_k4_d1000_seed0/config.json +337 -0
- 05_token_interface/multitoken_n2_k4_d1000_seed0/run_info.json +13 -0
- 05_token_interface/multitoken_n2_k4_d100_seed0/config.json +333 -0
- 05_token_interface/multitoken_n2_k4_d100_seed0/run_info.json +13 -0
- 05_token_interface/multitoken_n2_k4_d2900_seed0/config.json +333 -0
- 05_token_interface/multitoken_n2_k4_d2900_seed0/run_info.json +13 -0
04_robocasa_scaling/README.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Section 4 — Multitask control and scaling
|
| 2 |
+
|
| 3 |
+
Paper: **Can WorldToken achieve effective multitask control and scale systematically?**
|
| 4 |
+
|
| 5 |
+
Fifteen final scaling-grid checkpoints (N1–N3, seed 0, K=1, C_train=10) and the D300 / seed-123 BC-Transformer baseline.
|
| 6 |
+
|
| 7 |
+
| Run | Training steps | Checkpoint file |
|
| 8 |
+
|---|---:|---|
|
| 9 |
+
| [scaling_n1_d50_seed0](scaling_n1_d50_seed0/) | 5,000 | `checkpoint_step_00005000.pt` |
|
| 10 |
+
| [scaling_n2_d50_seed0](scaling_n2_d50_seed0/) | 5,000 | `checkpoint_step_00005000.pt` |
|
| 11 |
+
| [scaling_n3_d50_seed0](scaling_n3_d50_seed0/) | 5,000 | `checkpoint_step_00005000.pt` |
|
| 12 |
+
| [scaling_n1_d100_seed0](scaling_n1_d100_seed0/) | 10,000 | `checkpoint_step_00010000.pt` |
|
| 13 |
+
| [scaling_n2_d100_seed0](scaling_n2_d100_seed0/) | 10,000 | `checkpoint_step_00010000.pt` |
|
| 14 |
+
| [scaling_n3_d100_seed0](scaling_n3_d100_seed0/) | 10,000 | `checkpoint_step_00010000.pt` |
|
| 15 |
+
| [bc_transformer_d300_seed123](bc_transformer_d300_seed123/) | 500,000 | `models/model_epoch_1000.pth` |
|
| 16 |
+
| [scaling_n1_d300_seed0](scaling_n1_d300_seed0/) | 30,000 | `checkpoint_step_00030000.pt` |
|
| 17 |
+
| [scaling_n2_d300_seed0](scaling_n2_d300_seed0/) | 30,000 | `checkpoint_step_00030000.pt` |
|
| 18 |
+
| [scaling_n3_d300_seed0](scaling_n3_d300_seed0/) | 30,000 | `checkpoint_step_00030000.pt` |
|
| 19 |
+
| [scaling_n1_d1000_seed0](scaling_n1_d1000_seed0/) | 100,000 | `checkpoint_step_00100000.pt` |
|
| 20 |
+
| [scaling_n2_d1000_seed0](scaling_n2_d1000_seed0/) | 100,000 | `checkpoint_step_00100000.pt` |
|
| 21 |
+
| [scaling_n3_d1000_seed0](scaling_n3_d1000_seed0/) | 100,000 | `checkpoint_step_00100000.pt` |
|
| 22 |
+
| [scaling_n1_d2900_seed0](scaling_n1_d2900_seed0/) | 280,000 | `checkpoint_step_00280000.pt` |
|
| 23 |
+
| [scaling_n2_d2900_seed0](scaling_n2_d2900_seed0/) | 280,000 | `checkpoint_step_00280000.pt` |
|
| 24 |
+
| [scaling_n3_d2900_seed0](scaling_n3_d2900_seed0/) | 280,000 | `checkpoint_step_00280000.pt` |
|
| 25 |
+
|
| 26 |
+
BC-Transformer ends at epoch 1,000, corresponding to 500,000 training steps.
|
04_robocasa_scaling/bc_transformer_d300_seed123/config.json
ADDED
|
@@ -0,0 +1,462 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"algo_name": "bc",
|
| 3 |
+
"experiment": {
|
| 4 |
+
"name": "e0_bc_xfmr_mg23_d300_seed123_official500k_cuda4",
|
| 5 |
+
"validate": false,
|
| 6 |
+
"logging": {
|
| 7 |
+
"terminal_output_to_txt": true,
|
| 8 |
+
"log_tb": true,
|
| 9 |
+
"log_wandb": false,
|
| 10 |
+
"wandb_proj_name": "robocasa_official_baselines_mg23_d300"
|
| 11 |
+
},
|
| 12 |
+
"mse": {
|
| 13 |
+
"enabled": false,
|
| 14 |
+
"every_n_epochs": 50,
|
| 15 |
+
"on_save_ckpt": true,
|
| 16 |
+
"num_samples": 20,
|
| 17 |
+
"visualize": true
|
| 18 |
+
},
|
| 19 |
+
"save": {
|
| 20 |
+
"enabled": true,
|
| 21 |
+
"every_n_seconds": null,
|
| 22 |
+
"every_n_epochs": 100,
|
| 23 |
+
"epochs": [],
|
| 24 |
+
"on_best_validation": false,
|
| 25 |
+
"on_best_rollout_return": false,
|
| 26 |
+
"on_best_rollout_success_rate": false
|
| 27 |
+
},
|
| 28 |
+
"epoch_every_n_steps": 500,
|
| 29 |
+
"validation_epoch_every_n_steps": 10,
|
| 30 |
+
"env": null,
|
| 31 |
+
"additional_envs": null,
|
| 32 |
+
"render": false,
|
| 33 |
+
"render_video": false,
|
| 34 |
+
"keep_all_videos": false,
|
| 35 |
+
"video_skip": 5,
|
| 36 |
+
"rollout": {
|
| 37 |
+
"enabled": false,
|
| 38 |
+
"n": 50,
|
| 39 |
+
"horizon": 500,
|
| 40 |
+
"rate": 100,
|
| 41 |
+
"warmstart": 0,
|
| 42 |
+
"terminate_on_success": true,
|
| 43 |
+
"batched": false,
|
| 44 |
+
"num_batch_envs": 5
|
| 45 |
+
},
|
| 46 |
+
"env_meta_update_dict": {
|
| 47 |
+
"env_kwargs": {
|
| 48 |
+
"generative_textures": null,
|
| 49 |
+
"scene_split": null,
|
| 50 |
+
"style_ids": null,
|
| 51 |
+
"layout_ids": null,
|
| 52 |
+
"layout_and_style_ids": [
|
| 53 |
+
[
|
| 54 |
+
1,
|
| 55 |
+
1
|
| 56 |
+
],
|
| 57 |
+
[
|
| 58 |
+
2,
|
| 59 |
+
2
|
| 60 |
+
],
|
| 61 |
+
[
|
| 62 |
+
4,
|
| 63 |
+
4
|
| 64 |
+
],
|
| 65 |
+
[
|
| 66 |
+
6,
|
| 67 |
+
9
|
| 68 |
+
],
|
| 69 |
+
[
|
| 70 |
+
7,
|
| 71 |
+
10
|
| 72 |
+
]
|
| 73 |
+
],
|
| 74 |
+
"randomize_cameras": false,
|
| 75 |
+
"obj_instance_split": "B"
|
| 76 |
+
}
|
| 77 |
+
},
|
| 78 |
+
"ckpt_path": null
|
| 79 |
+
},
|
| 80 |
+
"train": {
|
| 81 |
+
"data": [
|
| 82 |
+
{
|
| 83 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 84 |
+
"horizon": 700,
|
| 85 |
+
"do_eval": false,
|
| 86 |
+
"filter_key": "300_demos"
|
| 87 |
+
},
|
| 88 |
+
{
|
| 89 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 90 |
+
"horizon": 500,
|
| 91 |
+
"do_eval": false,
|
| 92 |
+
"filter_key": "300_demos"
|
| 93 |
+
},
|
| 94 |
+
{
|
| 95 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 96 |
+
"horizon": 500,
|
| 97 |
+
"do_eval": false,
|
| 98 |
+
"filter_key": "300_demos"
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 102 |
+
"horizon": 300,
|
| 103 |
+
"do_eval": false,
|
| 104 |
+
"filter_key": "300_demos"
|
| 105 |
+
},
|
| 106 |
+
{
|
| 107 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 108 |
+
"horizon": 600,
|
| 109 |
+
"do_eval": false,
|
| 110 |
+
"filter_key": "300_demos"
|
| 111 |
+
},
|
| 112 |
+
{
|
| 113 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 114 |
+
"horizon": 600,
|
| 115 |
+
"do_eval": false,
|
| 116 |
+
"filter_key": "300_demos"
|
| 117 |
+
},
|
| 118 |
+
{
|
| 119 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 120 |
+
"horizon": 500,
|
| 121 |
+
"do_eval": false,
|
| 122 |
+
"filter_key": "300_demos"
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 126 |
+
"horizon": 500,
|
| 127 |
+
"do_eval": false,
|
| 128 |
+
"filter_key": "300_demos"
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"horizon": 500,
|
| 133 |
+
"do_eval": false,
|
| 134 |
+
"filter_key": "300_demos"
|
| 135 |
+
},
|
| 136 |
+
{
|
| 137 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"horizon": 500,
|
| 139 |
+
"do_eval": true,
|
| 140 |
+
"filter_key": "300_demos"
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"horizon": 600,
|
| 145 |
+
"do_eval": false,
|
| 146 |
+
"filter_key": "300_demos"
|
| 147 |
+
},
|
| 148 |
+
{
|
| 149 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"horizon": 700,
|
| 151 |
+
"do_eval": true,
|
| 152 |
+
"filter_key": "300_demos"
|
| 153 |
+
},
|
| 154 |
+
{
|
| 155 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 156 |
+
"horizon": 500,
|
| 157 |
+
"do_eval": false,
|
| 158 |
+
"filter_key": "300_demos"
|
| 159 |
+
},
|
| 160 |
+
{
|
| 161 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 162 |
+
"horizon": 500,
|
| 163 |
+
"do_eval": false,
|
| 164 |
+
"filter_key": "300_demos"
|
| 165 |
+
},
|
| 166 |
+
{
|
| 167 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 168 |
+
"horizon": 500,
|
| 169 |
+
"do_eval": false,
|
| 170 |
+
"filter_key": "300_demos"
|
| 171 |
+
},
|
| 172 |
+
{
|
| 173 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 174 |
+
"horizon": 500,
|
| 175 |
+
"do_eval": false,
|
| 176 |
+
"filter_key": "300_demos"
|
| 177 |
+
},
|
| 178 |
+
{
|
| 179 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 180 |
+
"horizon": 500,
|
| 181 |
+
"do_eval": false,
|
| 182 |
+
"filter_key": "300_demos"
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 186 |
+
"horizon": 500,
|
| 187 |
+
"do_eval": false,
|
| 188 |
+
"filter_key": "300_demos"
|
| 189 |
+
},
|
| 190 |
+
{
|
| 191 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 192 |
+
"horizon": 500,
|
| 193 |
+
"do_eval": false,
|
| 194 |
+
"filter_key": "300_demos"
|
| 195 |
+
},
|
| 196 |
+
{
|
| 197 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 198 |
+
"horizon": 500,
|
| 199 |
+
"do_eval": false,
|
| 200 |
+
"filter_key": "300_demos"
|
| 201 |
+
},
|
| 202 |
+
{
|
| 203 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 204 |
+
"horizon": 500,
|
| 205 |
+
"do_eval": false,
|
| 206 |
+
"filter_key": "300_demos"
|
| 207 |
+
},
|
| 208 |
+
{
|
| 209 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 210 |
+
"horizon": 500,
|
| 211 |
+
"do_eval": false,
|
| 212 |
+
"filter_key": "300_demos"
|
| 213 |
+
},
|
| 214 |
+
{
|
| 215 |
+
"path": "${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 216 |
+
"horizon": 500,
|
| 217 |
+
"do_eval": false,
|
| 218 |
+
"filter_key": "300_demos"
|
| 219 |
+
}
|
| 220 |
+
],
|
| 221 |
+
"output_dir": "${EXPERIMENT_ROOT}/E0/runs",
|
| 222 |
+
"num_data_workers": 5,
|
| 223 |
+
"hdf5_cache_mode": null,
|
| 224 |
+
"hdf5_use_swmr": true,
|
| 225 |
+
"hdf5_load_next_obs": false,
|
| 226 |
+
"hdf5_normalize_obs": false,
|
| 227 |
+
"hdf5_filter_key": null,
|
| 228 |
+
"hdf5_validation_filter_key": null,
|
| 229 |
+
"seq_length": 10,
|
| 230 |
+
"pad_seq_length": true,
|
| 231 |
+
"frame_stack": 10,
|
| 232 |
+
"pad_frame_stack": true,
|
| 233 |
+
"dataset_keys": [
|
| 234 |
+
"actions"
|
| 235 |
+
],
|
| 236 |
+
"action_keys": [
|
| 237 |
+
"actions"
|
| 238 |
+
],
|
| 239 |
+
"action_config": {
|
| 240 |
+
"actions": {
|
| 241 |
+
"normalization": null
|
| 242 |
+
},
|
| 243 |
+
"actions_abs": {
|
| 244 |
+
"normalization": "min_max"
|
| 245 |
+
},
|
| 246 |
+
"action_dict/abs_pos": {
|
| 247 |
+
"normalization": "min_max"
|
| 248 |
+
},
|
| 249 |
+
"action_dict/abs_rot_axis_angle": {
|
| 250 |
+
"normalization": "min_max",
|
| 251 |
+
"format": "rot_axis_angle"
|
| 252 |
+
},
|
| 253 |
+
"action_dict/abs_rot_6d": {
|
| 254 |
+
"normalization": null,
|
| 255 |
+
"format": "rot_6d"
|
| 256 |
+
},
|
| 257 |
+
"action_dict/rel_pos": {
|
| 258 |
+
"normalization": null
|
| 259 |
+
},
|
| 260 |
+
"action_dict/rel_rot_axis_angle": {
|
| 261 |
+
"normalization": null,
|
| 262 |
+
"format": "rot_axis_angle"
|
| 263 |
+
},
|
| 264 |
+
"action_dict/rel_rot_6d": {
|
| 265 |
+
"normalization": null,
|
| 266 |
+
"format": "rot_6d"
|
| 267 |
+
},
|
| 268 |
+
"action_dict/gripper": {
|
| 269 |
+
"normalization": null
|
| 270 |
+
},
|
| 271 |
+
"action_dict/base_mode": {
|
| 272 |
+
"normalization": null
|
| 273 |
+
}
|
| 274 |
+
},
|
| 275 |
+
"goal_mode": null,
|
| 276 |
+
"cuda": true,
|
| 277 |
+
"batch_size": 16,
|
| 278 |
+
"num_epochs": 1000,
|
| 279 |
+
"seed": 123,
|
| 280 |
+
"max_grad_norm": 100.0,
|
| 281 |
+
"data_format": "robomimic",
|
| 282 |
+
"shuffled_obs_key_groups": null
|
| 283 |
+
},
|
| 284 |
+
"algo": {
|
| 285 |
+
"optim_params": {
|
| 286 |
+
"policy": {
|
| 287 |
+
"optimizer_type": "adamw",
|
| 288 |
+
"learning_rate": {
|
| 289 |
+
"initial": 0.0001,
|
| 290 |
+
"decay_factor": 1.0,
|
| 291 |
+
"epoch_schedule": [
|
| 292 |
+
100
|
| 293 |
+
],
|
| 294 |
+
"scheduler_type": "constant_with_warmup"
|
| 295 |
+
},
|
| 296 |
+
"regularization": {
|
| 297 |
+
"L2": 0.01
|
| 298 |
+
}
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"loss": {
|
| 302 |
+
"l2_weight": 1.0,
|
| 303 |
+
"l1_weight": 0.0,
|
| 304 |
+
"cos_weight": 0.0
|
| 305 |
+
},
|
| 306 |
+
"actor_layer_dims": [],
|
| 307 |
+
"gaussian": {
|
| 308 |
+
"enabled": false,
|
| 309 |
+
"fixed_std": false,
|
| 310 |
+
"init_std": 0.1,
|
| 311 |
+
"min_std": 0.01,
|
| 312 |
+
"std_activation": "softplus",
|
| 313 |
+
"low_noise_eval": true
|
| 314 |
+
},
|
| 315 |
+
"gmm": {
|
| 316 |
+
"enabled": true,
|
| 317 |
+
"num_modes": 5,
|
| 318 |
+
"min_std": 0.005,
|
| 319 |
+
"std_activation": "softplus",
|
| 320 |
+
"low_noise_eval": true
|
| 321 |
+
},
|
| 322 |
+
"vae": {
|
| 323 |
+
"enabled": false,
|
| 324 |
+
"latent_dim": 14,
|
| 325 |
+
"latent_clip": null,
|
| 326 |
+
"kl_weight": 1.0,
|
| 327 |
+
"decoder": {
|
| 328 |
+
"is_conditioned": true,
|
| 329 |
+
"reconstruction_sum_across_elements": false
|
| 330 |
+
},
|
| 331 |
+
"prior": {
|
| 332 |
+
"learn": false,
|
| 333 |
+
"is_conditioned": false,
|
| 334 |
+
"use_gmm": false,
|
| 335 |
+
"gmm_num_modes": 10,
|
| 336 |
+
"gmm_learn_weights": false,
|
| 337 |
+
"use_categorical": false,
|
| 338 |
+
"categorical_dim": 10,
|
| 339 |
+
"categorical_gumbel_softmax_hard": false,
|
| 340 |
+
"categorical_init_temp": 1.0,
|
| 341 |
+
"categorical_temp_anneal_step": 0.001,
|
| 342 |
+
"categorical_min_temp": 0.3
|
| 343 |
+
},
|
| 344 |
+
"encoder_layer_dims": [
|
| 345 |
+
300,
|
| 346 |
+
400
|
| 347 |
+
],
|
| 348 |
+
"decoder_layer_dims": [
|
| 349 |
+
300,
|
| 350 |
+
400
|
| 351 |
+
],
|
| 352 |
+
"prior_layer_dims": [
|
| 353 |
+
300,
|
| 354 |
+
400
|
| 355 |
+
]
|
| 356 |
+
},
|
| 357 |
+
"rnn": {
|
| 358 |
+
"enabled": false,
|
| 359 |
+
"horizon": 10,
|
| 360 |
+
"hidden_dim": 400,
|
| 361 |
+
"rnn_type": "LSTM",
|
| 362 |
+
"num_layers": 2,
|
| 363 |
+
"open_loop": false,
|
| 364 |
+
"kwargs": {
|
| 365 |
+
"bidirectional": false
|
| 366 |
+
}
|
| 367 |
+
},
|
| 368 |
+
"transformer": {
|
| 369 |
+
"enabled": true,
|
| 370 |
+
"context_length": 10,
|
| 371 |
+
"embed_dim": 512,
|
| 372 |
+
"num_layers": 6,
|
| 373 |
+
"num_heads": 8,
|
| 374 |
+
"emb_dropout": 0.1,
|
| 375 |
+
"attn_dropout": 0.1,
|
| 376 |
+
"block_output_dropout": 0.1,
|
| 377 |
+
"sinusoidal_embedding": false,
|
| 378 |
+
"activation": "gelu",
|
| 379 |
+
"supervise_all_steps": true,
|
| 380 |
+
"nn_parameter_for_timesteps": true,
|
| 381 |
+
"pred_future_acs": true,
|
| 382 |
+
"causal": false
|
| 383 |
+
},
|
| 384 |
+
"language_conditioned": true
|
| 385 |
+
},
|
| 386 |
+
"observation": {
|
| 387 |
+
"modalities": {
|
| 388 |
+
"obs": {
|
| 389 |
+
"low_dim": [
|
| 390 |
+
"robot0_base_to_eef_pos",
|
| 391 |
+
"robot0_base_to_eef_quat",
|
| 392 |
+
"robot0_base_pos",
|
| 393 |
+
"robot0_base_quat",
|
| 394 |
+
"robot0_gripper_qpos"
|
| 395 |
+
],
|
| 396 |
+
"rgb": [
|
| 397 |
+
"robot0_agentview_left_image",
|
| 398 |
+
"robot0_agentview_right_image",
|
| 399 |
+
"robot0_eye_in_hand_image"
|
| 400 |
+
],
|
| 401 |
+
"depth": [],
|
| 402 |
+
"scan": []
|
| 403 |
+
},
|
| 404 |
+
"goal": {
|
| 405 |
+
"low_dim": [],
|
| 406 |
+
"rgb": [],
|
| 407 |
+
"depth": [],
|
| 408 |
+
"scan": []
|
| 409 |
+
}
|
| 410 |
+
},
|
| 411 |
+
"encoder": {
|
| 412 |
+
"low_dim": {
|
| 413 |
+
"core_class": null,
|
| 414 |
+
"core_kwargs": {},
|
| 415 |
+
"obs_randomizer_class": null,
|
| 416 |
+
"obs_randomizer_kwargs": {}
|
| 417 |
+
},
|
| 418 |
+
"rgb": {
|
| 419 |
+
"core_class": "VisualCoreLanguageConditioned",
|
| 420 |
+
"core_kwargs": {
|
| 421 |
+
"feature_dimension": 64,
|
| 422 |
+
"backbone_class": "ResNet18ConvFiLM",
|
| 423 |
+
"backbone_kwargs": {
|
| 424 |
+
"pretrained": false,
|
| 425 |
+
"input_coord_conv": false
|
| 426 |
+
},
|
| 427 |
+
"pool_class": "SpatialSoftmax",
|
| 428 |
+
"pool_kwargs": {
|
| 429 |
+
"num_kp": 32,
|
| 430 |
+
"learnable_temperature": false,
|
| 431 |
+
"temperature": 1.0,
|
| 432 |
+
"noise_std": 0.0
|
| 433 |
+
}
|
| 434 |
+
},
|
| 435 |
+
"obs_randomizer_class": "CropRandomizer",
|
| 436 |
+
"obs_randomizer_kwargs": {
|
| 437 |
+
"crop_height": 116,
|
| 438 |
+
"crop_width": 116,
|
| 439 |
+
"num_crops": 1,
|
| 440 |
+
"pos_enc": false
|
| 441 |
+
}
|
| 442 |
+
},
|
| 443 |
+
"depth": {
|
| 444 |
+
"core_class": "VisualCore",
|
| 445 |
+
"core_kwargs": {},
|
| 446 |
+
"obs_randomizer_class": null,
|
| 447 |
+
"obs_randomizer_kwargs": {}
|
| 448 |
+
},
|
| 449 |
+
"scan": {
|
| 450 |
+
"core_class": "ScanCore",
|
| 451 |
+
"core_kwargs": {},
|
| 452 |
+
"obs_randomizer_class": null,
|
| 453 |
+
"obs_randomizer_kwargs": {}
|
| 454 |
+
}
|
| 455 |
+
}
|
| 456 |
+
},
|
| 457 |
+
"meta": {
|
| 458 |
+
"hp_base_config_file": null,
|
| 459 |
+
"hp_keys": [],
|
| 460 |
+
"hp_values": []
|
| 461 |
+
}
|
| 462 |
+
}
|
04_robocasa_scaling/bc_transformer_d300_seed123/run_info.json
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "bc_transformer_d300_seed123",
|
| 3 |
+
"original_run_name": "e0_bc_xfmr_mg23_d300_seed123_official500k_cuda4",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 300,
|
| 7 |
+
"C_train": 10,
|
| 8 |
+
"seed": 123
|
| 9 |
+
},
|
| 10 |
+
"final_training_step": 500000
|
| 11 |
+
}
|
04_robocasa_scaling/scaling_n1_d1000_seed0/checkpoint_step_00100000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:adf58c8a3cf22a7733968c3c2d42e407fad079d4d7b747d186ae80098e742693
|
| 3 |
+
size 531280901
|
04_robocasa_scaling/scaling_n1_d1000_seed0/config.json
ADDED
|
@@ -0,0 +1,329 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"d_model": 128,
|
| 7 |
+
"denoising_steps": 20,
|
| 8 |
+
"dropout": 0.0,
|
| 9 |
+
"h_adaln_bottleneck": false,
|
| 10 |
+
"n_heads": 4,
|
| 11 |
+
"n_layers": 1,
|
| 12 |
+
"use_h_cross_attn": false,
|
| 13 |
+
"use_obs_cross_attn": false
|
| 14 |
+
},
|
| 15 |
+
"type": "diffusion_dit"
|
| 16 |
+
},
|
| 17 |
+
"action_head_lr": 0.0003,
|
| 18 |
+
"action_spec": {
|
| 19 |
+
"dim": 12,
|
| 20 |
+
"discrete_dims": [
|
| 21 |
+
6,
|
| 22 |
+
11
|
| 23 |
+
]
|
| 24 |
+
},
|
| 25 |
+
"action_weight": 1.0,
|
| 26 |
+
"adamw_beta1": 0.9,
|
| 27 |
+
"adamw_beta2": 0.95,
|
| 28 |
+
"adamw_eps": 1e-08,
|
| 29 |
+
"allow_holdout_overlap": false,
|
| 30 |
+
"attn_backend": "auto",
|
| 31 |
+
"attn_impl": "eager",
|
| 32 |
+
"backbone_lr": null,
|
| 33 |
+
"backbone_type": "qwen2",
|
| 34 |
+
"batch_size": 48,
|
| 35 |
+
"cnn_depth": 48,
|
| 36 |
+
"cnn_kernel": 5,
|
| 37 |
+
"cnn_mults": [
|
| 38 |
+
2,
|
| 39 |
+
3,
|
| 40 |
+
4,
|
| 41 |
+
4
|
| 42 |
+
],
|
| 43 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_d1000_n1_world2/phase1grid_d1000_n1_nodyn_seed0_100k.yaml",
|
| 44 |
+
"crops_per_demo": 8,
|
| 45 |
+
"d_model": 2048,
|
| 46 |
+
"dataset": [
|
| 47 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 48 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 70 |
+
],
|
| 71 |
+
"ddp_timeout_minutes": 180.0,
|
| 72 |
+
"denoising_steps": 20,
|
| 73 |
+
"device": "cuda",
|
| 74 |
+
"diffusion_mlp_dims": [
|
| 75 |
+
1024,
|
| 76 |
+
1024,
|
| 77 |
+
1024
|
| 78 |
+
],
|
| 79 |
+
"diffusion_residual_style": true,
|
| 80 |
+
"diffusion_time_dim": 16,
|
| 81 |
+
"dropout": 0.1,
|
| 82 |
+
"dry_run_data": false,
|
| 83 |
+
"dynamics": {
|
| 84 |
+
"enabled": false,
|
| 85 |
+
"params": {},
|
| 86 |
+
"type": "none"
|
| 87 |
+
},
|
| 88 |
+
"dynamics_decoder_type": "film",
|
| 89 |
+
"enable_pred_next": false,
|
| 90 |
+
"encoder": {
|
| 91 |
+
"params": {
|
| 92 |
+
"cnn_depth": 48,
|
| 93 |
+
"cnn_kernel": 5,
|
| 94 |
+
"cnn_mults": [
|
| 95 |
+
2,
|
| 96 |
+
3,
|
| 97 |
+
4,
|
| 98 |
+
8,
|
| 99 |
+
16
|
| 100 |
+
],
|
| 101 |
+
"d_model": 512,
|
| 102 |
+
"dropout": 0.0,
|
| 103 |
+
"mlp_ratio": 4,
|
| 104 |
+
"n_fusion_layers": 1,
|
| 105 |
+
"n_heads": 4,
|
| 106 |
+
"readout_depth": 1,
|
| 107 |
+
"readout_mlp_ratio": 2,
|
| 108 |
+
"readout_queries": 4,
|
| 109 |
+
"rope_theta": 10000.0,
|
| 110 |
+
"use_proprio": true
|
| 111 |
+
},
|
| 112 |
+
"type": "attn_fusion_latent_token"
|
| 113 |
+
},
|
| 114 |
+
"encoder_lr": 0.0006,
|
| 115 |
+
"eval_at_start": false,
|
| 116 |
+
"eval_batch_size": 64,
|
| 117 |
+
"eval_crops_per_demo": 8,
|
| 118 |
+
"eval_every": 5000,
|
| 119 |
+
"eval_num_workers": 24,
|
| 120 |
+
"eval_seed": 0,
|
| 121 |
+
"ffn_hidden_size": 1024,
|
| 122 |
+
"filter_key": "wl23_mg1000_seed0",
|
| 123 |
+
"grad_accum_steps": 2,
|
| 124 |
+
"grad_clip": 1.0,
|
| 125 |
+
"hdf5_paths": [
|
| 126 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 127 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 149 |
+
],
|
| 150 |
+
"holdout_action_mean_samples": 1,
|
| 151 |
+
"holdout_crops_per_demo": 8,
|
| 152 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 153 |
+
"holdout_per_task_metrics": true,
|
| 154 |
+
"holdout_rmse_samplers": [
|
| 155 |
+
"deterministic",
|
| 156 |
+
"stochastic"
|
| 157 |
+
],
|
| 158 |
+
"holdout_sample_deterministic": null,
|
| 159 |
+
"holdout_sample_prefix_horizon": 4,
|
| 160 |
+
"holdout_sample_rmse": true,
|
| 161 |
+
"holdout_size": 2300,
|
| 162 |
+
"image_emb_dim": 512,
|
| 163 |
+
"init_device": "meta",
|
| 164 |
+
"input_norm": true,
|
| 165 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 166 |
+
"lang_emb_mode": "clip",
|
| 167 |
+
"lang_encoder_device": "auto",
|
| 168 |
+
"lang_obs_emb_dim": 256,
|
| 169 |
+
"log_every": 25,
|
| 170 |
+
"lora": false,
|
| 171 |
+
"lora_alpha": 32,
|
| 172 |
+
"lora_dropout": 0.0,
|
| 173 |
+
"lora_rank": 16,
|
| 174 |
+
"lora_target": [
|
| 175 |
+
"q_proj",
|
| 176 |
+
"k_proj",
|
| 177 |
+
"v_proj",
|
| 178 |
+
"o_proj",
|
| 179 |
+
"gate_proj",
|
| 180 |
+
"up_proj",
|
| 181 |
+
"down_proj"
|
| 182 |
+
],
|
| 183 |
+
"lr": 0.0003,
|
| 184 |
+
"max_context_len": 1024,
|
| 185 |
+
"max_steps": 100000,
|
| 186 |
+
"min_lr_ratio": 0.1,
|
| 187 |
+
"model": {
|
| 188 |
+
"action_chunk_len": 10,
|
| 189 |
+
"latent_dim": 512
|
| 190 |
+
},
|
| 191 |
+
"model_dtype": "float32",
|
| 192 |
+
"n_heads": 16,
|
| 193 |
+
"n_kv_heads": null,
|
| 194 |
+
"n_layers": 2,
|
| 195 |
+
"norm_bias_weight_decay": null,
|
| 196 |
+
"norm_fit_batches": 0,
|
| 197 |
+
"num_workers": 24,
|
| 198 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 199 |
+
"obs_spec": {
|
| 200 |
+
"image_channels": 3,
|
| 201 |
+
"image_hw": [
|
| 202 |
+
128,
|
| 203 |
+
128
|
| 204 |
+
],
|
| 205 |
+
"image_keys": [
|
| 206 |
+
"robot0_agentview_left_image",
|
| 207 |
+
"robot0_agentview_right_image",
|
| 208 |
+
"robot0_eye_in_hand_image"
|
| 209 |
+
],
|
| 210 |
+
"lang_dim": 768,
|
| 211 |
+
"layout": "NHWC",
|
| 212 |
+
"low_dim_dims": [
|
| 213 |
+
3,
|
| 214 |
+
4,
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
2
|
| 218 |
+
],
|
| 219 |
+
"low_dim_keys": [
|
| 220 |
+
"robot0_base_to_eef_pos",
|
| 221 |
+
"robot0_base_to_eef_quat",
|
| 222 |
+
"robot0_base_pos",
|
| 223 |
+
"robot0_base_quat",
|
| 224 |
+
"robot0_gripper_qpos"
|
| 225 |
+
]
|
| 226 |
+
},
|
| 227 |
+
"obs_stride": 4,
|
| 228 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n1_d1000_seed0",
|
| 229 |
+
"persistent_workers": true,
|
| 230 |
+
"precision": "bf16",
|
| 231 |
+
"pred_next_image_key_weights": {},
|
| 232 |
+
"pred_next_image_weight": 1.0,
|
| 233 |
+
"pred_next_lang_weight": 0.1,
|
| 234 |
+
"pred_next_mode": "terminal",
|
| 235 |
+
"pred_next_obs_offset": 1,
|
| 236 |
+
"pred_next_proprio_weight": 0.15,
|
| 237 |
+
"pred_next_steps": 4,
|
| 238 |
+
"pred_warmup_steps": 0,
|
| 239 |
+
"predictor_lr": 0.0006,
|
| 240 |
+
"prefetch_factor": 4,
|
| 241 |
+
"prepare_lang_cache_only": false,
|
| 242 |
+
"proprio_emb_dim": 128,
|
| 243 |
+
"recon": {
|
| 244 |
+
"enabled": false,
|
| 245 |
+
"params": {
|
| 246 |
+
"depth": 48,
|
| 247 |
+
"kernel_size": 5,
|
| 248 |
+
"mults": [
|
| 249 |
+
2,
|
| 250 |
+
3,
|
| 251 |
+
4,
|
| 252 |
+
4
|
| 253 |
+
]
|
| 254 |
+
}
|
| 255 |
+
},
|
| 256 |
+
"recon_image_weight": 1.0,
|
| 257 |
+
"recon_lang_weight": 0.1,
|
| 258 |
+
"recon_obs": false,
|
| 259 |
+
"recon_proprio_weight": 1.0,
|
| 260 |
+
"resume": null,
|
| 261 |
+
"resume_final_lr": null,
|
| 262 |
+
"resume_lr_schedule": "checkpoint",
|
| 263 |
+
"robomimic_src": null,
|
| 264 |
+
"save_every": 10000,
|
| 265 |
+
"save_holdout_traces": true,
|
| 266 |
+
"save_latest_checkpoint": true,
|
| 267 |
+
"scaling_action_head_d_model": 128,
|
| 268 |
+
"scaling_action_head_layers": 1,
|
| 269 |
+
"scaling_action_head_lr": 0.0003,
|
| 270 |
+
"scaling_backbone_d_model": 512,
|
| 271 |
+
"scaling_backbone_layers": 2,
|
| 272 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 273 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 274 |
+
"scaling_encoder_d_model": 512,
|
| 275 |
+
"scaling_encoder_layers": 1,
|
| 276 |
+
"scaling_encoder_lr": 0.0006,
|
| 277 |
+
"scaling_expected_world_size": 2,
|
| 278 |
+
"scaling_global_batch_size": 192,
|
| 279 |
+
"scaling_group_b_lr_role": null,
|
| 280 |
+
"scaling_latent_dim": 512,
|
| 281 |
+
"scaling_lr_action_head": 0.0003,
|
| 282 |
+
"scaling_lr_base": null,
|
| 283 |
+
"scaling_lr_base_frozen": false,
|
| 284 |
+
"scaling_lr_encoder": 0.0006,
|
| 285 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 286 |
+
"scaling_lr_lookup_calibration": false,
|
| 287 |
+
"scaling_lr_lookup_model": "n1",
|
| 288 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 289 |
+
"scaling_lr_predictor": 0.0006,
|
| 290 |
+
"scaling_lr_reference": null,
|
| 291 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 292 |
+
"scaling_lr_table_version": "n_lookup_v20260711_n5",
|
| 293 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 294 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 295 |
+
"scaling_per_rank_microbatch": 48,
|
| 296 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 297 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 298 |
+
"scaling_predictor_lr": 0.0006,
|
| 299 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 300 |
+
"scaling_recipe": "nodyn",
|
| 301 |
+
"scaling_run_name": "phase1grid_d1000_n1_nodyn_seed0_100k",
|
| 302 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 303 |
+
"scaling_suite": "phase1_grid",
|
| 304 |
+
"seed": 0,
|
| 305 |
+
"seq_len": 10,
|
| 306 |
+
"sequence_model": {
|
| 307 |
+
"backbone_type": "qwen2",
|
| 308 |
+
"hidden_dim": 512,
|
| 309 |
+
"params": {
|
| 310 |
+
"attn_impl": "sdpa",
|
| 311 |
+
"dropout": 0.1,
|
| 312 |
+
"ffn_hidden_size": 1408,
|
| 313 |
+
"input_norm": false,
|
| 314 |
+
"max_context_len": 1024,
|
| 315 |
+
"model_dtype": "float32",
|
| 316 |
+
"n_heads": 4,
|
| 317 |
+
"n_kv_heads": null,
|
| 318 |
+
"n_layers": 2
|
| 319 |
+
},
|
| 320 |
+
"type": "continuous_transformer"
|
| 321 |
+
},
|
| 322 |
+
"use_proprio": true,
|
| 323 |
+
"warmup_steps": 5000,
|
| 324 |
+
"weight_decay": 0.0,
|
| 325 |
+
"z_bottleneck": {
|
| 326 |
+
"dim": null
|
| 327 |
+
},
|
| 328 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 329 |
+
}
|
04_robocasa_scaling/scaling_n1_d1000_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n1_d1000_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d1000_n1_nodyn_seed0_100k_cuda4-5",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 1000,
|
| 7 |
+
"N": 1,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 100000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n1_d100_seed0/checkpoint_step_00010000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:22c14fc8c7dac158b75f7d3c2c8b3f545461f4486b6979a6c2644abdb79a7b14
|
| 3 |
+
size 531280901
|
04_robocasa_scaling/scaling_n1_d100_seed0/config.json
ADDED
|
@@ -0,0 +1,329 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"d_model": 128,
|
| 7 |
+
"denoising_steps": 20,
|
| 8 |
+
"dropout": 0.0,
|
| 9 |
+
"h_adaln_bottleneck": false,
|
| 10 |
+
"n_heads": 4,
|
| 11 |
+
"n_layers": 1,
|
| 12 |
+
"use_h_cross_attn": false,
|
| 13 |
+
"use_obs_cross_attn": false
|
| 14 |
+
},
|
| 15 |
+
"type": "diffusion_dit"
|
| 16 |
+
},
|
| 17 |
+
"action_head_lr": 0.0003,
|
| 18 |
+
"action_spec": {
|
| 19 |
+
"dim": 12,
|
| 20 |
+
"discrete_dims": [
|
| 21 |
+
6,
|
| 22 |
+
11
|
| 23 |
+
]
|
| 24 |
+
},
|
| 25 |
+
"action_weight": 1.0,
|
| 26 |
+
"adamw_beta1": 0.9,
|
| 27 |
+
"adamw_beta2": 0.95,
|
| 28 |
+
"adamw_eps": 1e-08,
|
| 29 |
+
"allow_holdout_overlap": false,
|
| 30 |
+
"attn_backend": "auto",
|
| 31 |
+
"attn_impl": "eager",
|
| 32 |
+
"backbone_lr": null,
|
| 33 |
+
"backbone_type": "qwen2",
|
| 34 |
+
"batch_size": 64,
|
| 35 |
+
"cnn_depth": 48,
|
| 36 |
+
"cnn_kernel": 5,
|
| 37 |
+
"cnn_mults": [
|
| 38 |
+
2,
|
| 39 |
+
3,
|
| 40 |
+
4,
|
| 41 |
+
4
|
| 42 |
+
],
|
| 43 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_d100_n1_world1/phase1grid_d100_n1_nodyn_seed0_10k.yaml",
|
| 44 |
+
"crops_per_demo": 8,
|
| 45 |
+
"d_model": 2048,
|
| 46 |
+
"dataset": [
|
| 47 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 48 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 70 |
+
],
|
| 71 |
+
"ddp_timeout_minutes": 180.0,
|
| 72 |
+
"denoising_steps": 20,
|
| 73 |
+
"device": "cuda",
|
| 74 |
+
"diffusion_mlp_dims": [
|
| 75 |
+
1024,
|
| 76 |
+
1024,
|
| 77 |
+
1024
|
| 78 |
+
],
|
| 79 |
+
"diffusion_residual_style": true,
|
| 80 |
+
"diffusion_time_dim": 16,
|
| 81 |
+
"dropout": 0.1,
|
| 82 |
+
"dry_run_data": false,
|
| 83 |
+
"dynamics": {
|
| 84 |
+
"enabled": false,
|
| 85 |
+
"params": {},
|
| 86 |
+
"type": "none"
|
| 87 |
+
},
|
| 88 |
+
"dynamics_decoder_type": "film",
|
| 89 |
+
"enable_pred_next": false,
|
| 90 |
+
"encoder": {
|
| 91 |
+
"params": {
|
| 92 |
+
"cnn_depth": 48,
|
| 93 |
+
"cnn_kernel": 5,
|
| 94 |
+
"cnn_mults": [
|
| 95 |
+
2,
|
| 96 |
+
3,
|
| 97 |
+
4,
|
| 98 |
+
8,
|
| 99 |
+
16
|
| 100 |
+
],
|
| 101 |
+
"d_model": 512,
|
| 102 |
+
"dropout": 0.0,
|
| 103 |
+
"mlp_ratio": 4,
|
| 104 |
+
"n_fusion_layers": 1,
|
| 105 |
+
"n_heads": 4,
|
| 106 |
+
"readout_depth": 1,
|
| 107 |
+
"readout_mlp_ratio": 2,
|
| 108 |
+
"readout_queries": 4,
|
| 109 |
+
"rope_theta": 10000.0,
|
| 110 |
+
"use_proprio": true
|
| 111 |
+
},
|
| 112 |
+
"type": "attn_fusion_latent_token"
|
| 113 |
+
},
|
| 114 |
+
"encoder_lr": 0.0006,
|
| 115 |
+
"eval_at_start": false,
|
| 116 |
+
"eval_batch_size": 64,
|
| 117 |
+
"eval_crops_per_demo": 8,
|
| 118 |
+
"eval_every": 500,
|
| 119 |
+
"eval_num_workers": 24,
|
| 120 |
+
"eval_seed": 0,
|
| 121 |
+
"ffn_hidden_size": 1024,
|
| 122 |
+
"filter_key": "wl23_mg100_seed0",
|
| 123 |
+
"grad_accum_steps": 3,
|
| 124 |
+
"grad_clip": 1.0,
|
| 125 |
+
"hdf5_paths": [
|
| 126 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 127 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 149 |
+
],
|
| 150 |
+
"holdout_action_mean_samples": 1,
|
| 151 |
+
"holdout_crops_per_demo": 8,
|
| 152 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 153 |
+
"holdout_per_task_metrics": true,
|
| 154 |
+
"holdout_rmse_samplers": [
|
| 155 |
+
"deterministic",
|
| 156 |
+
"stochastic"
|
| 157 |
+
],
|
| 158 |
+
"holdout_sample_deterministic": null,
|
| 159 |
+
"holdout_sample_prefix_horizon": 4,
|
| 160 |
+
"holdout_sample_rmse": true,
|
| 161 |
+
"holdout_size": 2300,
|
| 162 |
+
"image_emb_dim": 512,
|
| 163 |
+
"init_device": "meta",
|
| 164 |
+
"input_norm": true,
|
| 165 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 166 |
+
"lang_emb_mode": "clip",
|
| 167 |
+
"lang_encoder_device": "auto",
|
| 168 |
+
"lang_obs_emb_dim": 256,
|
| 169 |
+
"log_every": 25,
|
| 170 |
+
"lora": false,
|
| 171 |
+
"lora_alpha": 32,
|
| 172 |
+
"lora_dropout": 0.0,
|
| 173 |
+
"lora_rank": 16,
|
| 174 |
+
"lora_target": [
|
| 175 |
+
"q_proj",
|
| 176 |
+
"k_proj",
|
| 177 |
+
"v_proj",
|
| 178 |
+
"o_proj",
|
| 179 |
+
"gate_proj",
|
| 180 |
+
"up_proj",
|
| 181 |
+
"down_proj"
|
| 182 |
+
],
|
| 183 |
+
"lr": 0.0003,
|
| 184 |
+
"max_context_len": 1024,
|
| 185 |
+
"max_steps": 10000,
|
| 186 |
+
"min_lr_ratio": 0.1,
|
| 187 |
+
"model": {
|
| 188 |
+
"action_chunk_len": 10,
|
| 189 |
+
"latent_dim": 512
|
| 190 |
+
},
|
| 191 |
+
"model_dtype": "float32",
|
| 192 |
+
"n_heads": 16,
|
| 193 |
+
"n_kv_heads": null,
|
| 194 |
+
"n_layers": 2,
|
| 195 |
+
"norm_bias_weight_decay": null,
|
| 196 |
+
"norm_fit_batches": 0,
|
| 197 |
+
"num_workers": 24,
|
| 198 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 199 |
+
"obs_spec": {
|
| 200 |
+
"image_channels": 3,
|
| 201 |
+
"image_hw": [
|
| 202 |
+
128,
|
| 203 |
+
128
|
| 204 |
+
],
|
| 205 |
+
"image_keys": [
|
| 206 |
+
"robot0_agentview_left_image",
|
| 207 |
+
"robot0_agentview_right_image",
|
| 208 |
+
"robot0_eye_in_hand_image"
|
| 209 |
+
],
|
| 210 |
+
"lang_dim": 768,
|
| 211 |
+
"layout": "NHWC",
|
| 212 |
+
"low_dim_dims": [
|
| 213 |
+
3,
|
| 214 |
+
4,
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
2
|
| 218 |
+
],
|
| 219 |
+
"low_dim_keys": [
|
| 220 |
+
"robot0_base_to_eef_pos",
|
| 221 |
+
"robot0_base_to_eef_quat",
|
| 222 |
+
"robot0_base_pos",
|
| 223 |
+
"robot0_base_quat",
|
| 224 |
+
"robot0_gripper_qpos"
|
| 225 |
+
]
|
| 226 |
+
},
|
| 227 |
+
"obs_stride": 4,
|
| 228 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n1_d100_seed0",
|
| 229 |
+
"persistent_workers": true,
|
| 230 |
+
"precision": "bf16",
|
| 231 |
+
"pred_next_image_key_weights": {},
|
| 232 |
+
"pred_next_image_weight": 1.0,
|
| 233 |
+
"pred_next_lang_weight": 0.1,
|
| 234 |
+
"pred_next_mode": "terminal",
|
| 235 |
+
"pred_next_obs_offset": 1,
|
| 236 |
+
"pred_next_proprio_weight": 0.15,
|
| 237 |
+
"pred_next_steps": 4,
|
| 238 |
+
"pred_warmup_steps": 0,
|
| 239 |
+
"predictor_lr": 0.0006,
|
| 240 |
+
"prefetch_factor": 4,
|
| 241 |
+
"prepare_lang_cache_only": false,
|
| 242 |
+
"proprio_emb_dim": 128,
|
| 243 |
+
"recon": {
|
| 244 |
+
"enabled": false,
|
| 245 |
+
"params": {
|
| 246 |
+
"depth": 48,
|
| 247 |
+
"kernel_size": 5,
|
| 248 |
+
"mults": [
|
| 249 |
+
2,
|
| 250 |
+
3,
|
| 251 |
+
4,
|
| 252 |
+
4
|
| 253 |
+
]
|
| 254 |
+
}
|
| 255 |
+
},
|
| 256 |
+
"recon_image_weight": 1.0,
|
| 257 |
+
"recon_lang_weight": 0.1,
|
| 258 |
+
"recon_obs": false,
|
| 259 |
+
"recon_proprio_weight": 1.0,
|
| 260 |
+
"resume": null,
|
| 261 |
+
"resume_final_lr": null,
|
| 262 |
+
"resume_lr_schedule": "checkpoint",
|
| 263 |
+
"robomimic_src": null,
|
| 264 |
+
"save_every": 1000,
|
| 265 |
+
"save_holdout_traces": true,
|
| 266 |
+
"save_latest_checkpoint": true,
|
| 267 |
+
"scaling_action_head_d_model": 128,
|
| 268 |
+
"scaling_action_head_layers": 1,
|
| 269 |
+
"scaling_action_head_lr": 0.0003,
|
| 270 |
+
"scaling_backbone_d_model": 512,
|
| 271 |
+
"scaling_backbone_layers": 2,
|
| 272 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 273 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 274 |
+
"scaling_encoder_d_model": 512,
|
| 275 |
+
"scaling_encoder_layers": 1,
|
| 276 |
+
"scaling_encoder_lr": 0.0006,
|
| 277 |
+
"scaling_expected_world_size": 1,
|
| 278 |
+
"scaling_global_batch_size": 192,
|
| 279 |
+
"scaling_group_b_lr_role": null,
|
| 280 |
+
"scaling_latent_dim": 512,
|
| 281 |
+
"scaling_lr_action_head": 0.0003,
|
| 282 |
+
"scaling_lr_base": null,
|
| 283 |
+
"scaling_lr_base_frozen": false,
|
| 284 |
+
"scaling_lr_encoder": 0.0006,
|
| 285 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 286 |
+
"scaling_lr_lookup_calibration": false,
|
| 287 |
+
"scaling_lr_lookup_model": "n1",
|
| 288 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 289 |
+
"scaling_lr_predictor": 0.0006,
|
| 290 |
+
"scaling_lr_reference": null,
|
| 291 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 292 |
+
"scaling_lr_table_version": "n_lookup_v20260711_n5",
|
| 293 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 294 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 295 |
+
"scaling_per_rank_microbatch": 64,
|
| 296 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 297 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 298 |
+
"scaling_predictor_lr": 0.0006,
|
| 299 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 300 |
+
"scaling_recipe": "nodyn",
|
| 301 |
+
"scaling_run_name": "phase1grid_d100_n1_nodyn_seed0_10k",
|
| 302 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 303 |
+
"scaling_suite": "phase1_grid",
|
| 304 |
+
"seed": 0,
|
| 305 |
+
"seq_len": 10,
|
| 306 |
+
"sequence_model": {
|
| 307 |
+
"backbone_type": "qwen2",
|
| 308 |
+
"hidden_dim": 512,
|
| 309 |
+
"params": {
|
| 310 |
+
"attn_impl": "sdpa",
|
| 311 |
+
"dropout": 0.1,
|
| 312 |
+
"ffn_hidden_size": 1408,
|
| 313 |
+
"input_norm": false,
|
| 314 |
+
"max_context_len": 1024,
|
| 315 |
+
"model_dtype": "float32",
|
| 316 |
+
"n_heads": 4,
|
| 317 |
+
"n_kv_heads": null,
|
| 318 |
+
"n_layers": 2
|
| 319 |
+
},
|
| 320 |
+
"type": "continuous_transformer"
|
| 321 |
+
},
|
| 322 |
+
"use_proprio": true,
|
| 323 |
+
"warmup_steps": 500,
|
| 324 |
+
"weight_decay": 0.0,
|
| 325 |
+
"z_bottleneck": {
|
| 326 |
+
"dim": null
|
| 327 |
+
},
|
| 328 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 329 |
+
}
|
04_robocasa_scaling/scaling_n1_d100_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n1_d100_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d100_n1_nodyn_seed0_10k_cuda6",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 100,
|
| 7 |
+
"N": 1,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 10000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n1_d2900_seed0/checkpoint_step_00280000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c8e8d7446b9b077368928f26d82dbe3b41da3b878ce97f17d6ce819380eff432
|
| 3 |
+
size 531280901
|
04_robocasa_scaling/scaling_n1_d2900_seed0/config.json
ADDED
|
@@ -0,0 +1,329 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"d_model": 128,
|
| 7 |
+
"denoising_steps": 20,
|
| 8 |
+
"dropout": 0.0,
|
| 9 |
+
"h_adaln_bottleneck": false,
|
| 10 |
+
"n_heads": 4,
|
| 11 |
+
"n_layers": 1,
|
| 12 |
+
"use_h_cross_attn": false,
|
| 13 |
+
"use_obs_cross_attn": false
|
| 14 |
+
},
|
| 15 |
+
"type": "diffusion_dit"
|
| 16 |
+
},
|
| 17 |
+
"action_head_lr": 0.0003,
|
| 18 |
+
"action_spec": {
|
| 19 |
+
"dim": 12,
|
| 20 |
+
"discrete_dims": [
|
| 21 |
+
6,
|
| 22 |
+
11
|
| 23 |
+
]
|
| 24 |
+
},
|
| 25 |
+
"action_weight": 1.0,
|
| 26 |
+
"adamw_beta1": 0.9,
|
| 27 |
+
"adamw_beta2": 0.95,
|
| 28 |
+
"adamw_eps": 1e-08,
|
| 29 |
+
"allow_holdout_overlap": false,
|
| 30 |
+
"attn_backend": "auto",
|
| 31 |
+
"attn_impl": "eager",
|
| 32 |
+
"backbone_lr": null,
|
| 33 |
+
"backbone_type": "qwen2",
|
| 34 |
+
"batch_size": 48,
|
| 35 |
+
"cnn_depth": 48,
|
| 36 |
+
"cnn_kernel": 5,
|
| 37 |
+
"cnn_mults": [
|
| 38 |
+
2,
|
| 39 |
+
3,
|
| 40 |
+
4,
|
| 41 |
+
4
|
| 42 |
+
],
|
| 43 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_d2900_n1_world4/phase1grid_d2900_n1_nodyn_seed0_280k.yaml",
|
| 44 |
+
"crops_per_demo": 8,
|
| 45 |
+
"d_model": 2048,
|
| 46 |
+
"dataset": [
|
| 47 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 48 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 70 |
+
],
|
| 71 |
+
"ddp_timeout_minutes": 180.0,
|
| 72 |
+
"denoising_steps": 20,
|
| 73 |
+
"device": "cuda",
|
| 74 |
+
"diffusion_mlp_dims": [
|
| 75 |
+
1024,
|
| 76 |
+
1024,
|
| 77 |
+
1024
|
| 78 |
+
],
|
| 79 |
+
"diffusion_residual_style": true,
|
| 80 |
+
"diffusion_time_dim": 16,
|
| 81 |
+
"dropout": 0.1,
|
| 82 |
+
"dry_run_data": false,
|
| 83 |
+
"dynamics": {
|
| 84 |
+
"enabled": false,
|
| 85 |
+
"params": {},
|
| 86 |
+
"type": "none"
|
| 87 |
+
},
|
| 88 |
+
"dynamics_decoder_type": "film",
|
| 89 |
+
"enable_pred_next": false,
|
| 90 |
+
"encoder": {
|
| 91 |
+
"params": {
|
| 92 |
+
"cnn_depth": 48,
|
| 93 |
+
"cnn_kernel": 5,
|
| 94 |
+
"cnn_mults": [
|
| 95 |
+
2,
|
| 96 |
+
3,
|
| 97 |
+
4,
|
| 98 |
+
8,
|
| 99 |
+
16
|
| 100 |
+
],
|
| 101 |
+
"d_model": 512,
|
| 102 |
+
"dropout": 0.0,
|
| 103 |
+
"mlp_ratio": 4,
|
| 104 |
+
"n_fusion_layers": 1,
|
| 105 |
+
"n_heads": 4,
|
| 106 |
+
"readout_depth": 1,
|
| 107 |
+
"readout_mlp_ratio": 2,
|
| 108 |
+
"readout_queries": 4,
|
| 109 |
+
"rope_theta": 10000.0,
|
| 110 |
+
"use_proprio": true
|
| 111 |
+
},
|
| 112 |
+
"type": "attn_fusion_latent_token"
|
| 113 |
+
},
|
| 114 |
+
"encoder_lr": 0.0006,
|
| 115 |
+
"eval_at_start": false,
|
| 116 |
+
"eval_batch_size": 64,
|
| 117 |
+
"eval_crops_per_demo": 8,
|
| 118 |
+
"eval_every": 14000,
|
| 119 |
+
"eval_num_workers": 24,
|
| 120 |
+
"eval_seed": 0,
|
| 121 |
+
"ffn_hidden_size": 1024,
|
| 122 |
+
"filter_key": "wl23_mg2900_seed0",
|
| 123 |
+
"grad_accum_steps": 1,
|
| 124 |
+
"grad_clip": 1.0,
|
| 125 |
+
"hdf5_paths": [
|
| 126 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 127 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 149 |
+
],
|
| 150 |
+
"holdout_action_mean_samples": 1,
|
| 151 |
+
"holdout_crops_per_demo": 8,
|
| 152 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 153 |
+
"holdout_per_task_metrics": true,
|
| 154 |
+
"holdout_rmse_samplers": [
|
| 155 |
+
"deterministic",
|
| 156 |
+
"stochastic"
|
| 157 |
+
],
|
| 158 |
+
"holdout_sample_deterministic": null,
|
| 159 |
+
"holdout_sample_prefix_horizon": 4,
|
| 160 |
+
"holdout_sample_rmse": true,
|
| 161 |
+
"holdout_size": 2300,
|
| 162 |
+
"image_emb_dim": 512,
|
| 163 |
+
"init_device": "meta",
|
| 164 |
+
"input_norm": true,
|
| 165 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 166 |
+
"lang_emb_mode": "clip",
|
| 167 |
+
"lang_encoder_device": "auto",
|
| 168 |
+
"lang_obs_emb_dim": 256,
|
| 169 |
+
"log_every": 25,
|
| 170 |
+
"lora": false,
|
| 171 |
+
"lora_alpha": 32,
|
| 172 |
+
"lora_dropout": 0.0,
|
| 173 |
+
"lora_rank": 16,
|
| 174 |
+
"lora_target": [
|
| 175 |
+
"q_proj",
|
| 176 |
+
"k_proj",
|
| 177 |
+
"v_proj",
|
| 178 |
+
"o_proj",
|
| 179 |
+
"gate_proj",
|
| 180 |
+
"up_proj",
|
| 181 |
+
"down_proj"
|
| 182 |
+
],
|
| 183 |
+
"lr": 0.0003,
|
| 184 |
+
"max_context_len": 1024,
|
| 185 |
+
"max_steps": 280000,
|
| 186 |
+
"min_lr_ratio": 0.1,
|
| 187 |
+
"model": {
|
| 188 |
+
"action_chunk_len": 10,
|
| 189 |
+
"latent_dim": 512
|
| 190 |
+
},
|
| 191 |
+
"model_dtype": "float32",
|
| 192 |
+
"n_heads": 16,
|
| 193 |
+
"n_kv_heads": null,
|
| 194 |
+
"n_layers": 2,
|
| 195 |
+
"norm_bias_weight_decay": null,
|
| 196 |
+
"norm_fit_batches": 0,
|
| 197 |
+
"num_workers": 24,
|
| 198 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 199 |
+
"obs_spec": {
|
| 200 |
+
"image_channels": 3,
|
| 201 |
+
"image_hw": [
|
| 202 |
+
128,
|
| 203 |
+
128
|
| 204 |
+
],
|
| 205 |
+
"image_keys": [
|
| 206 |
+
"robot0_agentview_left_image",
|
| 207 |
+
"robot0_agentview_right_image",
|
| 208 |
+
"robot0_eye_in_hand_image"
|
| 209 |
+
],
|
| 210 |
+
"lang_dim": 768,
|
| 211 |
+
"layout": "NHWC",
|
| 212 |
+
"low_dim_dims": [
|
| 213 |
+
3,
|
| 214 |
+
4,
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
2
|
| 218 |
+
],
|
| 219 |
+
"low_dim_keys": [
|
| 220 |
+
"robot0_base_to_eef_pos",
|
| 221 |
+
"robot0_base_to_eef_quat",
|
| 222 |
+
"robot0_base_pos",
|
| 223 |
+
"robot0_base_quat",
|
| 224 |
+
"robot0_gripper_qpos"
|
| 225 |
+
]
|
| 226 |
+
},
|
| 227 |
+
"obs_stride": 4,
|
| 228 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n1_d2900_seed0",
|
| 229 |
+
"persistent_workers": true,
|
| 230 |
+
"precision": "bf16",
|
| 231 |
+
"pred_next_image_key_weights": {},
|
| 232 |
+
"pred_next_image_weight": 1.0,
|
| 233 |
+
"pred_next_lang_weight": 0.1,
|
| 234 |
+
"pred_next_mode": "terminal",
|
| 235 |
+
"pred_next_obs_offset": 1,
|
| 236 |
+
"pred_next_proprio_weight": 0.15,
|
| 237 |
+
"pred_next_steps": 4,
|
| 238 |
+
"pred_warmup_steps": 0,
|
| 239 |
+
"predictor_lr": 0.0006,
|
| 240 |
+
"prefetch_factor": 4,
|
| 241 |
+
"prepare_lang_cache_only": false,
|
| 242 |
+
"proprio_emb_dim": 128,
|
| 243 |
+
"recon": {
|
| 244 |
+
"enabled": false,
|
| 245 |
+
"params": {
|
| 246 |
+
"depth": 48,
|
| 247 |
+
"kernel_size": 5,
|
| 248 |
+
"mults": [
|
| 249 |
+
2,
|
| 250 |
+
3,
|
| 251 |
+
4,
|
| 252 |
+
4
|
| 253 |
+
]
|
| 254 |
+
}
|
| 255 |
+
},
|
| 256 |
+
"recon_image_weight": 1.0,
|
| 257 |
+
"recon_lang_weight": 0.1,
|
| 258 |
+
"recon_obs": false,
|
| 259 |
+
"recon_proprio_weight": 1.0,
|
| 260 |
+
"resume": null,
|
| 261 |
+
"resume_final_lr": null,
|
| 262 |
+
"resume_lr_schedule": "checkpoint",
|
| 263 |
+
"robomimic_src": null,
|
| 264 |
+
"save_every": 28000,
|
| 265 |
+
"save_holdout_traces": true,
|
| 266 |
+
"save_latest_checkpoint": true,
|
| 267 |
+
"scaling_action_head_d_model": 128,
|
| 268 |
+
"scaling_action_head_layers": 1,
|
| 269 |
+
"scaling_action_head_lr": 0.0003,
|
| 270 |
+
"scaling_backbone_d_model": 512,
|
| 271 |
+
"scaling_backbone_layers": 2,
|
| 272 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 273 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 274 |
+
"scaling_encoder_d_model": 512,
|
| 275 |
+
"scaling_encoder_layers": 1,
|
| 276 |
+
"scaling_encoder_lr": 0.0006,
|
| 277 |
+
"scaling_expected_world_size": 4,
|
| 278 |
+
"scaling_global_batch_size": 192,
|
| 279 |
+
"scaling_group_b_lr_role": null,
|
| 280 |
+
"scaling_latent_dim": 512,
|
| 281 |
+
"scaling_lr_action_head": 0.0003,
|
| 282 |
+
"scaling_lr_base": null,
|
| 283 |
+
"scaling_lr_base_frozen": false,
|
| 284 |
+
"scaling_lr_encoder": 0.0006,
|
| 285 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 286 |
+
"scaling_lr_lookup_calibration": false,
|
| 287 |
+
"scaling_lr_lookup_model": "n1",
|
| 288 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 289 |
+
"scaling_lr_predictor": 0.0006,
|
| 290 |
+
"scaling_lr_reference": null,
|
| 291 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 292 |
+
"scaling_lr_table_version": "n_lookup_v20260711_n5",
|
| 293 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 294 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 295 |
+
"scaling_per_rank_microbatch": 48,
|
| 296 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 297 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 298 |
+
"scaling_predictor_lr": 0.0006,
|
| 299 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 300 |
+
"scaling_recipe": "nodyn",
|
| 301 |
+
"scaling_run_name": "phase1grid_d2900_n1_nodyn_seed0_280k",
|
| 302 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 303 |
+
"scaling_suite": "phase1_grid",
|
| 304 |
+
"seed": 0,
|
| 305 |
+
"seq_len": 10,
|
| 306 |
+
"sequence_model": {
|
| 307 |
+
"backbone_type": "qwen2",
|
| 308 |
+
"hidden_dim": 512,
|
| 309 |
+
"params": {
|
| 310 |
+
"attn_impl": "sdpa",
|
| 311 |
+
"dropout": 0.1,
|
| 312 |
+
"ffn_hidden_size": 1408,
|
| 313 |
+
"input_norm": false,
|
| 314 |
+
"max_context_len": 1024,
|
| 315 |
+
"model_dtype": "float32",
|
| 316 |
+
"n_heads": 4,
|
| 317 |
+
"n_kv_heads": null,
|
| 318 |
+
"n_layers": 2
|
| 319 |
+
},
|
| 320 |
+
"type": "continuous_transformer"
|
| 321 |
+
},
|
| 322 |
+
"use_proprio": true,
|
| 323 |
+
"warmup_steps": 14000,
|
| 324 |
+
"weight_decay": 0.0,
|
| 325 |
+
"z_bottleneck": {
|
| 326 |
+
"dim": null
|
| 327 |
+
},
|
| 328 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 329 |
+
}
|
04_robocasa_scaling/scaling_n1_d2900_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n1_d2900_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d2900_n1_nodyn_seed0_280k_cuda0-1-2-3",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 2900,
|
| 7 |
+
"N": 1,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 280000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n1_d300_seed0/checkpoint_step_00030000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c3c340b937db28c148aeb9f4a161a2f039b86c8916fd95214b7bfaf72052fb7d
|
| 3 |
+
size 531280965
|
04_robocasa_scaling/scaling_n1_d300_seed0/config.json
ADDED
|
@@ -0,0 +1,331 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 128,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 512,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 4,
|
| 13 |
+
"n_layers": 1,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_d300_n1_world2/phase1grid_d300_n1_nodyn_seed0_30k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 512,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 1,
|
| 107 |
+
"n_heads": 4,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.0006,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 1500,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "300_demos",
|
| 125 |
+
"grad_accum_steps": 2,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 30000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 512
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n1_d300_seed0",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.0006,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 3000,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_d_model": 128,
|
| 270 |
+
"scaling_action_head_layers": 1,
|
| 271 |
+
"scaling_action_head_lr": 0.0003,
|
| 272 |
+
"scaling_backbone_d_model": 512,
|
| 273 |
+
"scaling_backbone_layers": 2,
|
| 274 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 275 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 276 |
+
"scaling_encoder_d_model": 512,
|
| 277 |
+
"scaling_encoder_layers": 1,
|
| 278 |
+
"scaling_encoder_lr": 0.0006,
|
| 279 |
+
"scaling_expected_world_size": 2,
|
| 280 |
+
"scaling_global_batch_size": 192,
|
| 281 |
+
"scaling_group_b_lr_role": null,
|
| 282 |
+
"scaling_latent_dim": 512,
|
| 283 |
+
"scaling_lr_action_head": 0.0003,
|
| 284 |
+
"scaling_lr_base": null,
|
| 285 |
+
"scaling_lr_base_frozen": false,
|
| 286 |
+
"scaling_lr_encoder": 0.0006,
|
| 287 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 288 |
+
"scaling_lr_lookup_calibration": false,
|
| 289 |
+
"scaling_lr_lookup_model": "n1",
|
| 290 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 291 |
+
"scaling_lr_predictor": 0.0006,
|
| 292 |
+
"scaling_lr_reference": null,
|
| 293 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 294 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 295 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 296 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 297 |
+
"scaling_per_rank_microbatch": 48,
|
| 298 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 299 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 300 |
+
"scaling_predictor_lr": 0.0006,
|
| 301 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 302 |
+
"scaling_recipe": "nodyn",
|
| 303 |
+
"scaling_run_name": "phase1grid_d300_n1_nodyn_seed0_30k",
|
| 304 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 305 |
+
"scaling_suite": "phase1_grid",
|
| 306 |
+
"seed": 0,
|
| 307 |
+
"seq_len": 10,
|
| 308 |
+
"sequence_model": {
|
| 309 |
+
"backbone_type": "qwen2",
|
| 310 |
+
"hidden_dim": 512,
|
| 311 |
+
"params": {
|
| 312 |
+
"attn_impl": "sdpa",
|
| 313 |
+
"dropout": 0.1,
|
| 314 |
+
"ffn_hidden_size": 1408,
|
| 315 |
+
"input_norm": false,
|
| 316 |
+
"max_context_len": 1024,
|
| 317 |
+
"model_dtype": "float32",
|
| 318 |
+
"n_heads": 4,
|
| 319 |
+
"n_kv_heads": null,
|
| 320 |
+
"n_layers": 2
|
| 321 |
+
},
|
| 322 |
+
"type": "continuous_transformer"
|
| 323 |
+
},
|
| 324 |
+
"use_proprio": true,
|
| 325 |
+
"warmup_steps": 1500,
|
| 326 |
+
"weight_decay": 0.0,
|
| 327 |
+
"z_bottleneck": {
|
| 328 |
+
"dim": null
|
| 329 |
+
},
|
| 330 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 331 |
+
}
|
04_robocasa_scaling/scaling_n1_d300_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n1_d300_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d300_n1_nodyn_seed0_30k_cuda2-4",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 300,
|
| 7 |
+
"N": 1,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 30000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n1_d50_seed0/checkpoint_step_00005000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:682682b296ebde324402a0edfaa296f9c13b1a912efd5ad130c48ac3a24ed9b9
|
| 3 |
+
size 531280837
|
04_robocasa_scaling/scaling_n1_d50_seed0/config.json
ADDED
|
@@ -0,0 +1,329 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"d_model": 128,
|
| 7 |
+
"denoising_steps": 20,
|
| 8 |
+
"dropout": 0.0,
|
| 9 |
+
"h_adaln_bottleneck": false,
|
| 10 |
+
"n_heads": 4,
|
| 11 |
+
"n_layers": 1,
|
| 12 |
+
"use_h_cross_attn": false,
|
| 13 |
+
"use_obs_cross_attn": false
|
| 14 |
+
},
|
| 15 |
+
"type": "diffusion_dit"
|
| 16 |
+
},
|
| 17 |
+
"action_head_lr": 0.0003,
|
| 18 |
+
"action_spec": {
|
| 19 |
+
"dim": 12,
|
| 20 |
+
"discrete_dims": [
|
| 21 |
+
6,
|
| 22 |
+
11
|
| 23 |
+
]
|
| 24 |
+
},
|
| 25 |
+
"action_weight": 1.0,
|
| 26 |
+
"adamw_beta1": 0.9,
|
| 27 |
+
"adamw_beta2": 0.95,
|
| 28 |
+
"adamw_eps": 1e-08,
|
| 29 |
+
"allow_holdout_overlap": false,
|
| 30 |
+
"attn_backend": "auto",
|
| 31 |
+
"attn_impl": "eager",
|
| 32 |
+
"backbone_lr": null,
|
| 33 |
+
"backbone_type": "qwen2",
|
| 34 |
+
"batch_size": 48,
|
| 35 |
+
"cnn_depth": 48,
|
| 36 |
+
"cnn_kernel": 5,
|
| 37 |
+
"cnn_mults": [
|
| 38 |
+
2,
|
| 39 |
+
3,
|
| 40 |
+
4,
|
| 41 |
+
4
|
| 42 |
+
],
|
| 43 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_d50_n1_world1/phase1grid_d50_n1_nodyn_seed0_5k.yaml",
|
| 44 |
+
"crops_per_demo": 8,
|
| 45 |
+
"d_model": 2048,
|
| 46 |
+
"dataset": [
|
| 47 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 48 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 70 |
+
],
|
| 71 |
+
"ddp_timeout_minutes": 180.0,
|
| 72 |
+
"denoising_steps": 20,
|
| 73 |
+
"device": "cuda",
|
| 74 |
+
"diffusion_mlp_dims": [
|
| 75 |
+
1024,
|
| 76 |
+
1024,
|
| 77 |
+
1024
|
| 78 |
+
],
|
| 79 |
+
"diffusion_residual_style": true,
|
| 80 |
+
"diffusion_time_dim": 16,
|
| 81 |
+
"dropout": 0.1,
|
| 82 |
+
"dry_run_data": false,
|
| 83 |
+
"dynamics": {
|
| 84 |
+
"enabled": false,
|
| 85 |
+
"params": {},
|
| 86 |
+
"type": "none"
|
| 87 |
+
},
|
| 88 |
+
"dynamics_decoder_type": "film",
|
| 89 |
+
"enable_pred_next": false,
|
| 90 |
+
"encoder": {
|
| 91 |
+
"params": {
|
| 92 |
+
"cnn_depth": 48,
|
| 93 |
+
"cnn_kernel": 5,
|
| 94 |
+
"cnn_mults": [
|
| 95 |
+
2,
|
| 96 |
+
3,
|
| 97 |
+
4,
|
| 98 |
+
8,
|
| 99 |
+
16
|
| 100 |
+
],
|
| 101 |
+
"d_model": 512,
|
| 102 |
+
"dropout": 0.0,
|
| 103 |
+
"mlp_ratio": 4,
|
| 104 |
+
"n_fusion_layers": 1,
|
| 105 |
+
"n_heads": 4,
|
| 106 |
+
"readout_depth": 1,
|
| 107 |
+
"readout_mlp_ratio": 2,
|
| 108 |
+
"readout_queries": 4,
|
| 109 |
+
"rope_theta": 10000.0,
|
| 110 |
+
"use_proprio": true
|
| 111 |
+
},
|
| 112 |
+
"type": "attn_fusion_latent_token"
|
| 113 |
+
},
|
| 114 |
+
"encoder_lr": 0.0006,
|
| 115 |
+
"eval_at_start": false,
|
| 116 |
+
"eval_batch_size": 64,
|
| 117 |
+
"eval_crops_per_demo": 8,
|
| 118 |
+
"eval_every": 250,
|
| 119 |
+
"eval_num_workers": 24,
|
| 120 |
+
"eval_seed": 0,
|
| 121 |
+
"ffn_hidden_size": 1024,
|
| 122 |
+
"filter_key": "wl23_mg50_seed0",
|
| 123 |
+
"grad_accum_steps": 4,
|
| 124 |
+
"grad_clip": 1.0,
|
| 125 |
+
"hdf5_paths": [
|
| 126 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 127 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 149 |
+
],
|
| 150 |
+
"holdout_action_mean_samples": 1,
|
| 151 |
+
"holdout_crops_per_demo": 8,
|
| 152 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 153 |
+
"holdout_per_task_metrics": true,
|
| 154 |
+
"holdout_rmse_samplers": [
|
| 155 |
+
"deterministic",
|
| 156 |
+
"stochastic"
|
| 157 |
+
],
|
| 158 |
+
"holdout_sample_deterministic": null,
|
| 159 |
+
"holdout_sample_prefix_horizon": 4,
|
| 160 |
+
"holdout_sample_rmse": true,
|
| 161 |
+
"holdout_size": 2300,
|
| 162 |
+
"image_emb_dim": 512,
|
| 163 |
+
"init_device": "meta",
|
| 164 |
+
"input_norm": true,
|
| 165 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 166 |
+
"lang_emb_mode": "clip",
|
| 167 |
+
"lang_encoder_device": "auto",
|
| 168 |
+
"lang_obs_emb_dim": 256,
|
| 169 |
+
"log_every": 25,
|
| 170 |
+
"lora": false,
|
| 171 |
+
"lora_alpha": 32,
|
| 172 |
+
"lora_dropout": 0.0,
|
| 173 |
+
"lora_rank": 16,
|
| 174 |
+
"lora_target": [
|
| 175 |
+
"q_proj",
|
| 176 |
+
"k_proj",
|
| 177 |
+
"v_proj",
|
| 178 |
+
"o_proj",
|
| 179 |
+
"gate_proj",
|
| 180 |
+
"up_proj",
|
| 181 |
+
"down_proj"
|
| 182 |
+
],
|
| 183 |
+
"lr": 0.0003,
|
| 184 |
+
"max_context_len": 1024,
|
| 185 |
+
"max_steps": 5000,
|
| 186 |
+
"min_lr_ratio": 0.1,
|
| 187 |
+
"model": {
|
| 188 |
+
"action_chunk_len": 10,
|
| 189 |
+
"latent_dim": 512
|
| 190 |
+
},
|
| 191 |
+
"model_dtype": "float32",
|
| 192 |
+
"n_heads": 16,
|
| 193 |
+
"n_kv_heads": null,
|
| 194 |
+
"n_layers": 2,
|
| 195 |
+
"norm_bias_weight_decay": null,
|
| 196 |
+
"norm_fit_batches": 0,
|
| 197 |
+
"num_workers": 24,
|
| 198 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 199 |
+
"obs_spec": {
|
| 200 |
+
"image_channels": 3,
|
| 201 |
+
"image_hw": [
|
| 202 |
+
128,
|
| 203 |
+
128
|
| 204 |
+
],
|
| 205 |
+
"image_keys": [
|
| 206 |
+
"robot0_agentview_left_image",
|
| 207 |
+
"robot0_agentview_right_image",
|
| 208 |
+
"robot0_eye_in_hand_image"
|
| 209 |
+
],
|
| 210 |
+
"lang_dim": 768,
|
| 211 |
+
"layout": "NHWC",
|
| 212 |
+
"low_dim_dims": [
|
| 213 |
+
3,
|
| 214 |
+
4,
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
2
|
| 218 |
+
],
|
| 219 |
+
"low_dim_keys": [
|
| 220 |
+
"robot0_base_to_eef_pos",
|
| 221 |
+
"robot0_base_to_eef_quat",
|
| 222 |
+
"robot0_base_pos",
|
| 223 |
+
"robot0_base_quat",
|
| 224 |
+
"robot0_gripper_qpos"
|
| 225 |
+
]
|
| 226 |
+
},
|
| 227 |
+
"obs_stride": 4,
|
| 228 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n1_d50_seed0",
|
| 229 |
+
"persistent_workers": true,
|
| 230 |
+
"precision": "bf16",
|
| 231 |
+
"pred_next_image_key_weights": {},
|
| 232 |
+
"pred_next_image_weight": 1.0,
|
| 233 |
+
"pred_next_lang_weight": 0.1,
|
| 234 |
+
"pred_next_mode": "terminal",
|
| 235 |
+
"pred_next_obs_offset": 1,
|
| 236 |
+
"pred_next_proprio_weight": 0.15,
|
| 237 |
+
"pred_next_steps": 4,
|
| 238 |
+
"pred_warmup_steps": 0,
|
| 239 |
+
"predictor_lr": 0.0006,
|
| 240 |
+
"prefetch_factor": 4,
|
| 241 |
+
"prepare_lang_cache_only": false,
|
| 242 |
+
"proprio_emb_dim": 128,
|
| 243 |
+
"recon": {
|
| 244 |
+
"enabled": false,
|
| 245 |
+
"params": {
|
| 246 |
+
"depth": 48,
|
| 247 |
+
"kernel_size": 5,
|
| 248 |
+
"mults": [
|
| 249 |
+
2,
|
| 250 |
+
3,
|
| 251 |
+
4,
|
| 252 |
+
4
|
| 253 |
+
]
|
| 254 |
+
}
|
| 255 |
+
},
|
| 256 |
+
"recon_image_weight": 1.0,
|
| 257 |
+
"recon_lang_weight": 0.1,
|
| 258 |
+
"recon_obs": false,
|
| 259 |
+
"recon_proprio_weight": 1.0,
|
| 260 |
+
"resume": null,
|
| 261 |
+
"resume_final_lr": null,
|
| 262 |
+
"resume_lr_schedule": "checkpoint",
|
| 263 |
+
"robomimic_src": null,
|
| 264 |
+
"save_every": 500,
|
| 265 |
+
"save_holdout_traces": true,
|
| 266 |
+
"save_latest_checkpoint": true,
|
| 267 |
+
"scaling_action_head_d_model": 128,
|
| 268 |
+
"scaling_action_head_layers": 1,
|
| 269 |
+
"scaling_action_head_lr": 0.0003,
|
| 270 |
+
"scaling_backbone_d_model": 512,
|
| 271 |
+
"scaling_backbone_layers": 2,
|
| 272 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 273 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 274 |
+
"scaling_encoder_d_model": 512,
|
| 275 |
+
"scaling_encoder_layers": 1,
|
| 276 |
+
"scaling_encoder_lr": 0.0006,
|
| 277 |
+
"scaling_expected_world_size": 1,
|
| 278 |
+
"scaling_global_batch_size": 192,
|
| 279 |
+
"scaling_group_b_lr_role": null,
|
| 280 |
+
"scaling_latent_dim": 512,
|
| 281 |
+
"scaling_lr_action_head": 0.0003,
|
| 282 |
+
"scaling_lr_base": null,
|
| 283 |
+
"scaling_lr_base_frozen": false,
|
| 284 |
+
"scaling_lr_encoder": 0.0006,
|
| 285 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 286 |
+
"scaling_lr_lookup_calibration": false,
|
| 287 |
+
"scaling_lr_lookup_model": "n1",
|
| 288 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 289 |
+
"scaling_lr_predictor": 0.0006,
|
| 290 |
+
"scaling_lr_reference": null,
|
| 291 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 292 |
+
"scaling_lr_table_version": "n_lookup_v20260711_n5",
|
| 293 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 294 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 295 |
+
"scaling_per_rank_microbatch": 48,
|
| 296 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 297 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 298 |
+
"scaling_predictor_lr": 0.0006,
|
| 299 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 300 |
+
"scaling_recipe": "nodyn",
|
| 301 |
+
"scaling_run_name": "phase1grid_d50_n1_nodyn_seed0_5k",
|
| 302 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 303 |
+
"scaling_suite": "phase1_grid",
|
| 304 |
+
"seed": 0,
|
| 305 |
+
"seq_len": 10,
|
| 306 |
+
"sequence_model": {
|
| 307 |
+
"backbone_type": "qwen2",
|
| 308 |
+
"hidden_dim": 512,
|
| 309 |
+
"params": {
|
| 310 |
+
"attn_impl": "sdpa",
|
| 311 |
+
"dropout": 0.1,
|
| 312 |
+
"ffn_hidden_size": 1408,
|
| 313 |
+
"input_norm": false,
|
| 314 |
+
"max_context_len": 1024,
|
| 315 |
+
"model_dtype": "float32",
|
| 316 |
+
"n_heads": 4,
|
| 317 |
+
"n_kv_heads": null,
|
| 318 |
+
"n_layers": 2
|
| 319 |
+
},
|
| 320 |
+
"type": "continuous_transformer"
|
| 321 |
+
},
|
| 322 |
+
"use_proprio": true,
|
| 323 |
+
"warmup_steps": 250,
|
| 324 |
+
"weight_decay": 0.0,
|
| 325 |
+
"z_bottleneck": {
|
| 326 |
+
"dim": null
|
| 327 |
+
},
|
| 328 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 329 |
+
}
|
04_robocasa_scaling/scaling_n1_d50_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n1_d50_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d50_n1_nodyn_seed0_5k_cuda6",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 50,
|
| 7 |
+
"N": 1,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 5000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n2_d1000_seed0/checkpoint_step_00100000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5371a75babaaf64880677ac3ed6db74f29dca9f2483e58998ecc6db7b1e6a158
|
| 3 |
+
size 1024355757
|
04_robocasa_scaling/scaling_n2_d1000_seed0/config.json
ADDED
|
@@ -0,0 +1,331 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 192,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 768,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 6,
|
| 13 |
+
"n_layers": 2,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_d1000_n2_world2/phase1grid_d1000_n2_nodyn_seed0_100k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 768,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 2,
|
| 107 |
+
"n_heads": 6,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.000425,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 5000,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "wl23_mg1000_seed0",
|
| 125 |
+
"grad_accum_steps": 2,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 100000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 768
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n2_d1000_seed0",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.000425,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 10000,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_d_model": 192,
|
| 270 |
+
"scaling_action_head_layers": 2,
|
| 271 |
+
"scaling_action_head_lr": 0.0003,
|
| 272 |
+
"scaling_backbone_d_model": 768,
|
| 273 |
+
"scaling_backbone_layers": 4,
|
| 274 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 275 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 276 |
+
"scaling_encoder_d_model": 768,
|
| 277 |
+
"scaling_encoder_layers": 2,
|
| 278 |
+
"scaling_encoder_lr": 0.000425,
|
| 279 |
+
"scaling_expected_world_size": 2,
|
| 280 |
+
"scaling_global_batch_size": 192,
|
| 281 |
+
"scaling_group_b_lr_role": null,
|
| 282 |
+
"scaling_latent_dim": 768,
|
| 283 |
+
"scaling_lr_action_head": 0.0003,
|
| 284 |
+
"scaling_lr_base": null,
|
| 285 |
+
"scaling_lr_base_frozen": false,
|
| 286 |
+
"scaling_lr_encoder": 0.000425,
|
| 287 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 288 |
+
"scaling_lr_lookup_calibration": false,
|
| 289 |
+
"scaling_lr_lookup_model": "n2",
|
| 290 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 291 |
+
"scaling_lr_predictor": 0.000425,
|
| 292 |
+
"scaling_lr_reference": null,
|
| 293 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 294 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 295 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 296 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 297 |
+
"scaling_per_rank_microbatch": 48,
|
| 298 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 299 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 300 |
+
"scaling_predictor_lr": 0.000425,
|
| 301 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 302 |
+
"scaling_recipe": "nodyn",
|
| 303 |
+
"scaling_run_name": "phase1grid_d1000_n2_nodyn_seed0_100k",
|
| 304 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 305 |
+
"scaling_suite": "phase1_grid",
|
| 306 |
+
"seed": 0,
|
| 307 |
+
"seq_len": 10,
|
| 308 |
+
"sequence_model": {
|
| 309 |
+
"backbone_type": "qwen2",
|
| 310 |
+
"hidden_dim": 768,
|
| 311 |
+
"params": {
|
| 312 |
+
"attn_impl": "sdpa",
|
| 313 |
+
"dropout": 0.1,
|
| 314 |
+
"ffn_hidden_size": 2048,
|
| 315 |
+
"input_norm": false,
|
| 316 |
+
"max_context_len": 1024,
|
| 317 |
+
"model_dtype": "float32",
|
| 318 |
+
"n_heads": 6,
|
| 319 |
+
"n_kv_heads": null,
|
| 320 |
+
"n_layers": 4
|
| 321 |
+
},
|
| 322 |
+
"type": "continuous_transformer"
|
| 323 |
+
},
|
| 324 |
+
"use_proprio": true,
|
| 325 |
+
"warmup_steps": 5000,
|
| 326 |
+
"weight_decay": 0.0,
|
| 327 |
+
"z_bottleneck": {
|
| 328 |
+
"dim": null
|
| 329 |
+
},
|
| 330 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 331 |
+
}
|
04_robocasa_scaling/scaling_n2_d1000_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n2_d1000_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d1000_n2_nodyn_seed0_100k_cuda4-5",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 1000,
|
| 7 |
+
"N": 2,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 100000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n2_d100_seed0/checkpoint_step_00010000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f03b059f1c4b471497a3b3b4678cfd8b5df585be82f21eefc21b34ab5a5b65c1
|
| 3 |
+
size 1024357869
|
04_robocasa_scaling/scaling_n2_d100_seed0/config.json
ADDED
|
@@ -0,0 +1,331 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 192,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 768,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 6,
|
| 13 |
+
"n_layers": 2,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_resume_n2_d100/phase1grid_d100_n2_nodyn_seed0_10k_resume_mb48_accum4.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 768,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 2,
|
| 107 |
+
"n_heads": 6,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.000425,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 500,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "wl23_mg100_seed0",
|
| 125 |
+
"grad_accum_steps": 4,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 10000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 768
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n2_d100_seed0",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.000425,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": "${CHECKPOINT_ROOT}/checkpoint_step_00008000.pt",
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 1000,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_d_model": 192,
|
| 270 |
+
"scaling_action_head_layers": 2,
|
| 271 |
+
"scaling_action_head_lr": 0.0003,
|
| 272 |
+
"scaling_backbone_d_model": 768,
|
| 273 |
+
"scaling_backbone_layers": 4,
|
| 274 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 275 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 276 |
+
"scaling_encoder_d_model": 768,
|
| 277 |
+
"scaling_encoder_layers": 2,
|
| 278 |
+
"scaling_encoder_lr": 0.000425,
|
| 279 |
+
"scaling_expected_world_size": 1,
|
| 280 |
+
"scaling_global_batch_size": 192,
|
| 281 |
+
"scaling_group_b_lr_role": null,
|
| 282 |
+
"scaling_latent_dim": 768,
|
| 283 |
+
"scaling_lr_action_head": 0.0003,
|
| 284 |
+
"scaling_lr_base": null,
|
| 285 |
+
"scaling_lr_base_frozen": false,
|
| 286 |
+
"scaling_lr_encoder": 0.000425,
|
| 287 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 288 |
+
"scaling_lr_lookup_calibration": false,
|
| 289 |
+
"scaling_lr_lookup_model": "n2",
|
| 290 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 291 |
+
"scaling_lr_predictor": 0.000425,
|
| 292 |
+
"scaling_lr_reference": null,
|
| 293 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 294 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 295 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 296 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 297 |
+
"scaling_per_rank_microbatch": 48,
|
| 298 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 299 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 300 |
+
"scaling_predictor_lr": 0.000425,
|
| 301 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 302 |
+
"scaling_recipe": "nodyn",
|
| 303 |
+
"scaling_run_name": "phase1grid_d100_n2_nodyn_seed0_10k",
|
| 304 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 305 |
+
"scaling_suite": "phase1_grid",
|
| 306 |
+
"seed": 0,
|
| 307 |
+
"seq_len": 10,
|
| 308 |
+
"sequence_model": {
|
| 309 |
+
"backbone_type": "qwen2",
|
| 310 |
+
"hidden_dim": 768,
|
| 311 |
+
"params": {
|
| 312 |
+
"attn_impl": "sdpa",
|
| 313 |
+
"dropout": 0.1,
|
| 314 |
+
"ffn_hidden_size": 2048,
|
| 315 |
+
"input_norm": false,
|
| 316 |
+
"max_context_len": 1024,
|
| 317 |
+
"model_dtype": "float32",
|
| 318 |
+
"n_heads": 6,
|
| 319 |
+
"n_kv_heads": null,
|
| 320 |
+
"n_layers": 4
|
| 321 |
+
},
|
| 322 |
+
"type": "continuous_transformer"
|
| 323 |
+
},
|
| 324 |
+
"use_proprio": true,
|
| 325 |
+
"warmup_steps": 500,
|
| 326 |
+
"weight_decay": 0.0,
|
| 327 |
+
"z_bottleneck": {
|
| 328 |
+
"dim": null
|
| 329 |
+
},
|
| 330 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 331 |
+
}
|
04_robocasa_scaling/scaling_n2_d100_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n2_d100_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d100_n2_nodyn_seed0_10k_cuda5",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 100,
|
| 7 |
+
"N": 2,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 10000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n2_d2900_seed0/checkpoint_step_00280000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d321278276a2f3919cc1c3b5472675ee3b7fffe32a44d2559dc1dddabae3b100
|
| 3 |
+
size 1024355757
|
04_robocasa_scaling/scaling_n2_d2900_seed0/config.json
ADDED
|
@@ -0,0 +1,331 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 192,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 768,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 6,
|
| 13 |
+
"n_layers": 2,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_d2900_n2_world4/phase1grid_d2900_n2_nodyn_seed0_280k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 768,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 2,
|
| 107 |
+
"n_heads": 6,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.000425,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 14000,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "wl23_mg2900_seed0",
|
| 125 |
+
"grad_accum_steps": 1,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 280000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 768
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n2_d2900_seed0",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.000425,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 28000,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_d_model": 192,
|
| 270 |
+
"scaling_action_head_layers": 2,
|
| 271 |
+
"scaling_action_head_lr": 0.0003,
|
| 272 |
+
"scaling_backbone_d_model": 768,
|
| 273 |
+
"scaling_backbone_layers": 4,
|
| 274 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 275 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 276 |
+
"scaling_encoder_d_model": 768,
|
| 277 |
+
"scaling_encoder_layers": 2,
|
| 278 |
+
"scaling_encoder_lr": 0.000425,
|
| 279 |
+
"scaling_expected_world_size": 4,
|
| 280 |
+
"scaling_global_batch_size": 192,
|
| 281 |
+
"scaling_group_b_lr_role": null,
|
| 282 |
+
"scaling_latent_dim": 768,
|
| 283 |
+
"scaling_lr_action_head": 0.0003,
|
| 284 |
+
"scaling_lr_base": null,
|
| 285 |
+
"scaling_lr_base_frozen": false,
|
| 286 |
+
"scaling_lr_encoder": 0.000425,
|
| 287 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 288 |
+
"scaling_lr_lookup_calibration": false,
|
| 289 |
+
"scaling_lr_lookup_model": "n2",
|
| 290 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 291 |
+
"scaling_lr_predictor": 0.000425,
|
| 292 |
+
"scaling_lr_reference": null,
|
| 293 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 294 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 295 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 296 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 297 |
+
"scaling_per_rank_microbatch": 48,
|
| 298 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 299 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 300 |
+
"scaling_predictor_lr": 0.000425,
|
| 301 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 302 |
+
"scaling_recipe": "nodyn",
|
| 303 |
+
"scaling_run_name": "phase1grid_d2900_n2_nodyn_seed0_280k",
|
| 304 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 305 |
+
"scaling_suite": "phase1_grid",
|
| 306 |
+
"seed": 0,
|
| 307 |
+
"seq_len": 10,
|
| 308 |
+
"sequence_model": {
|
| 309 |
+
"backbone_type": "qwen2",
|
| 310 |
+
"hidden_dim": 768,
|
| 311 |
+
"params": {
|
| 312 |
+
"attn_impl": "sdpa",
|
| 313 |
+
"dropout": 0.1,
|
| 314 |
+
"ffn_hidden_size": 2048,
|
| 315 |
+
"input_norm": false,
|
| 316 |
+
"max_context_len": 1024,
|
| 317 |
+
"model_dtype": "float32",
|
| 318 |
+
"n_heads": 6,
|
| 319 |
+
"n_kv_heads": null,
|
| 320 |
+
"n_layers": 4
|
| 321 |
+
},
|
| 322 |
+
"type": "continuous_transformer"
|
| 323 |
+
},
|
| 324 |
+
"use_proprio": true,
|
| 325 |
+
"warmup_steps": 14000,
|
| 326 |
+
"weight_decay": 0.0,
|
| 327 |
+
"z_bottleneck": {
|
| 328 |
+
"dim": null
|
| 329 |
+
},
|
| 330 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 331 |
+
}
|
04_robocasa_scaling/scaling_n2_d2900_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n2_d2900_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d2900_n2_nodyn_seed0_280k_cuda1-2-4-5",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 2900,
|
| 7 |
+
"N": 2,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 280000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n2_d300_seed0/checkpoint_step_00030000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c209e131e5438c4430d7274f764e1867023101d6a58b71bae7c233405920d8d6
|
| 3 |
+
size 1024355757
|
04_robocasa_scaling/scaling_n2_d300_seed0/config.json
ADDED
|
@@ -0,0 +1,331 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 192,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 768,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 6,
|
| 13 |
+
"n_layers": 2,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_d300_n2_world2/phase1grid_d300_n2_nodyn_seed0_30k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 768,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 2,
|
| 107 |
+
"n_heads": 6,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.000425,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 1500,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "300_demos",
|
| 125 |
+
"grad_accum_steps": 2,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 30000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 768
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n2_d300_seed0",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.000425,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 3000,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_d_model": 192,
|
| 270 |
+
"scaling_action_head_layers": 2,
|
| 271 |
+
"scaling_action_head_lr": 0.0003,
|
| 272 |
+
"scaling_backbone_d_model": 768,
|
| 273 |
+
"scaling_backbone_layers": 4,
|
| 274 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 275 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 276 |
+
"scaling_encoder_d_model": 768,
|
| 277 |
+
"scaling_encoder_layers": 2,
|
| 278 |
+
"scaling_encoder_lr": 0.000425,
|
| 279 |
+
"scaling_expected_world_size": 2,
|
| 280 |
+
"scaling_global_batch_size": 192,
|
| 281 |
+
"scaling_group_b_lr_role": null,
|
| 282 |
+
"scaling_latent_dim": 768,
|
| 283 |
+
"scaling_lr_action_head": 0.0003,
|
| 284 |
+
"scaling_lr_base": null,
|
| 285 |
+
"scaling_lr_base_frozen": false,
|
| 286 |
+
"scaling_lr_encoder": 0.000425,
|
| 287 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 288 |
+
"scaling_lr_lookup_calibration": false,
|
| 289 |
+
"scaling_lr_lookup_model": "n2",
|
| 290 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 291 |
+
"scaling_lr_predictor": 0.000425,
|
| 292 |
+
"scaling_lr_reference": null,
|
| 293 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 294 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 295 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 296 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 297 |
+
"scaling_per_rank_microbatch": 48,
|
| 298 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 299 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 300 |
+
"scaling_predictor_lr": 0.000425,
|
| 301 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 302 |
+
"scaling_recipe": "nodyn",
|
| 303 |
+
"scaling_run_name": "phase1grid_d300_n2_nodyn_seed0_30k",
|
| 304 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 305 |
+
"scaling_suite": "phase1_grid",
|
| 306 |
+
"seed": 0,
|
| 307 |
+
"seq_len": 10,
|
| 308 |
+
"sequence_model": {
|
| 309 |
+
"backbone_type": "qwen2",
|
| 310 |
+
"hidden_dim": 768,
|
| 311 |
+
"params": {
|
| 312 |
+
"attn_impl": "sdpa",
|
| 313 |
+
"dropout": 0.1,
|
| 314 |
+
"ffn_hidden_size": 2048,
|
| 315 |
+
"input_norm": false,
|
| 316 |
+
"max_context_len": 1024,
|
| 317 |
+
"model_dtype": "float32",
|
| 318 |
+
"n_heads": 6,
|
| 319 |
+
"n_kv_heads": null,
|
| 320 |
+
"n_layers": 4
|
| 321 |
+
},
|
| 322 |
+
"type": "continuous_transformer"
|
| 323 |
+
},
|
| 324 |
+
"use_proprio": true,
|
| 325 |
+
"warmup_steps": 1500,
|
| 326 |
+
"weight_decay": 0.0,
|
| 327 |
+
"z_bottleneck": {
|
| 328 |
+
"dim": null
|
| 329 |
+
},
|
| 330 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 331 |
+
}
|
04_robocasa_scaling/scaling_n2_d300_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n2_d300_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d300_n2_nodyn_seed0_30k_cuda0-1",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 300,
|
| 7 |
+
"N": 2,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 30000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n2_d50_seed0/checkpoint_step_00005000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:68a50cdaab591898585aac9470797ee73fc08a0fec036ac0bb038d839fbaa9fe
|
| 3 |
+
size 1024355757
|
04_robocasa_scaling/scaling_n2_d50_seed0/config.json
ADDED
|
@@ -0,0 +1,331 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 192,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 768,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 6,
|
| 13 |
+
"n_layers": 2,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/Phase1_5x5_wd0_newfamily/configs_d50_d100_n2_world1/phase1grid_d50_n2_nodyn_seed0_5k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 768,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 2,
|
| 107 |
+
"n_heads": 6,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.000425,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 250,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "wl23_mg50_seed0",
|
| 125 |
+
"grad_accum_steps": 4,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 5000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 768
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${RELEASE_ROOT}/04_robocasa_scaling/scaling_n2_d50_seed0",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.000425,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 500,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_d_model": 192,
|
| 270 |
+
"scaling_action_head_layers": 2,
|
| 271 |
+
"scaling_action_head_lr": 0.0003,
|
| 272 |
+
"scaling_backbone_d_model": 768,
|
| 273 |
+
"scaling_backbone_layers": 4,
|
| 274 |
+
"scaling_batch_policy": "phase1_grid_h100_v20260711",
|
| 275 |
+
"scaling_batch_policy_d50_microbatch": 48,
|
| 276 |
+
"scaling_encoder_d_model": 768,
|
| 277 |
+
"scaling_encoder_layers": 2,
|
| 278 |
+
"scaling_encoder_lr": 0.000425,
|
| 279 |
+
"scaling_expected_world_size": 1,
|
| 280 |
+
"scaling_global_batch_size": 192,
|
| 281 |
+
"scaling_group_b_lr_role": null,
|
| 282 |
+
"scaling_latent_dim": 768,
|
| 283 |
+
"scaling_lr_action_head": 0.0003,
|
| 284 |
+
"scaling_lr_base": null,
|
| 285 |
+
"scaling_lr_base_frozen": false,
|
| 286 |
+
"scaling_lr_encoder": 0.000425,
|
| 287 |
+
"scaling_lr_formula_size_axis": "discrete_family_lookup",
|
| 288 |
+
"scaling_lr_lookup_calibration": false,
|
| 289 |
+
"scaling_lr_lookup_model": "n2",
|
| 290 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 291 |
+
"scaling_lr_predictor": 0.000425,
|
| 292 |
+
"scaling_lr_reference": null,
|
| 293 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 294 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 295 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 296 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 297 |
+
"scaling_per_rank_microbatch": 48,
|
| 298 |
+
"scaling_phase1_grid_seed_policy": "one_training_seed_per_cell_seed0",
|
| 299 |
+
"scaling_phase1_topology": "phase1_5x5_seed0_v20260711",
|
| 300 |
+
"scaling_predictor_lr": 0.000425,
|
| 301 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 302 |
+
"scaling_recipe": "nodyn",
|
| 303 |
+
"scaling_run_name": "phase1grid_d50_n2_nodyn_seed0_5k",
|
| 304 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/prefix04",
|
| 305 |
+
"scaling_suite": "phase1_grid",
|
| 306 |
+
"seed": 0,
|
| 307 |
+
"seq_len": 10,
|
| 308 |
+
"sequence_model": {
|
| 309 |
+
"backbone_type": "qwen2",
|
| 310 |
+
"hidden_dim": 768,
|
| 311 |
+
"params": {
|
| 312 |
+
"attn_impl": "sdpa",
|
| 313 |
+
"dropout": 0.1,
|
| 314 |
+
"ffn_hidden_size": 2048,
|
| 315 |
+
"input_norm": false,
|
| 316 |
+
"max_context_len": 1024,
|
| 317 |
+
"model_dtype": "float32",
|
| 318 |
+
"n_heads": 6,
|
| 319 |
+
"n_kv_heads": null,
|
| 320 |
+
"n_layers": 4
|
| 321 |
+
},
|
| 322 |
+
"type": "continuous_transformer"
|
| 323 |
+
},
|
| 324 |
+
"use_proprio": true,
|
| 325 |
+
"warmup_steps": 250,
|
| 326 |
+
"weight_decay": 0.0,
|
| 327 |
+
"z_bottleneck": {
|
| 328 |
+
"dim": null
|
| 329 |
+
},
|
| 330 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_02.json.gz"
|
| 331 |
+
}
|
04_robocasa_scaling/scaling_n2_d50_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n2_d50_seed0",
|
| 3 |
+
"original_run_name": "phase1grid_d50_n2_nodyn_seed0_5k_cuda4",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 50,
|
| 7 |
+
"N": 2,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 5000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n3_d1000_seed0/config.json
ADDED
|
@@ -0,0 +1,327 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 256,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 1024,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 8,
|
| 13 |
+
"n_layers": 4,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/E1_scaling_grid/configs_world4/e1_grid_d1000_n3_nodyn_seed0_100k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 1024,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 4,
|
| 107 |
+
"n_heads": 8,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.0003,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 5000,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "wl23_mg1000_seed0",
|
| 125 |
+
"grad_accum_steps": 1,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 100000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 1024
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${EXPERIMENT_ROOT}/E1_scaling_grid/runs/e1_grid_d1000_n3_nodyn_seed0_100k_cuda0-4-5-7",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.0003,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 10000,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_layers": 4,
|
| 270 |
+
"scaling_action_head_width": 256,
|
| 271 |
+
"scaling_analysis_suites": [
|
| 272 |
+
"e1_grid"
|
| 273 |
+
],
|
| 274 |
+
"scaling_backbone_layers": 8,
|
| 275 |
+
"scaling_encoder_layers": 4,
|
| 276 |
+
"scaling_expected_world_size": 4,
|
| 277 |
+
"scaling_family_width": 1024,
|
| 278 |
+
"scaling_global_batch_size": 192,
|
| 279 |
+
"scaling_grid_data_demos_per_task": 1000,
|
| 280 |
+
"scaling_grid_model_size": "n3",
|
| 281 |
+
"scaling_grid_protocol": "direct_5x5_seed0_first_pass_v20260711",
|
| 282 |
+
"scaling_lr_action_head": 0.0003,
|
| 283 |
+
"scaling_lr_encoder": 0.0003,
|
| 284 |
+
"scaling_lr_lookup_model": "n3",
|
| 285 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 286 |
+
"scaling_lr_param_groups": [
|
| 287 |
+
"encoder",
|
| 288 |
+
"predictor",
|
| 289 |
+
"action_head"
|
| 290 |
+
],
|
| 291 |
+
"scaling_lr_predictor": 0.0003,
|
| 292 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 293 |
+
"scaling_lr_table_version": "n_lookup_v20260712",
|
| 294 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 295 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 296 |
+
"scaling_per_rank_microbatch": 48,
|
| 297 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 298 |
+
"scaling_recipe": "nodyn",
|
| 299 |
+
"scaling_run_name": "e1_grid_d1000_n3_nodyn_seed0_100k",
|
| 300 |
+
"scaling_secondary_holdout_metric": "action_rmse",
|
| 301 |
+
"scaling_suite": "e1_grid",
|
| 302 |
+
"seed": 0,
|
| 303 |
+
"seq_len": 10,
|
| 304 |
+
"sequence_model": {
|
| 305 |
+
"backbone_type": "qwen2",
|
| 306 |
+
"hidden_dim": 1024,
|
| 307 |
+
"params": {
|
| 308 |
+
"attn_impl": "sdpa",
|
| 309 |
+
"dropout": 0.1,
|
| 310 |
+
"ffn_hidden_size": 2816,
|
| 311 |
+
"input_norm": false,
|
| 312 |
+
"max_context_len": 1024,
|
| 313 |
+
"model_dtype": "float32",
|
| 314 |
+
"n_heads": 8,
|
| 315 |
+
"n_kv_heads": null,
|
| 316 |
+
"n_layers": 8
|
| 317 |
+
},
|
| 318 |
+
"type": "continuous_transformer"
|
| 319 |
+
},
|
| 320 |
+
"use_proprio": true,
|
| 321 |
+
"warmup_steps": 5000,
|
| 322 |
+
"weight_decay": 0.0,
|
| 323 |
+
"z_bottleneck": {
|
| 324 |
+
"dim": null
|
| 325 |
+
},
|
| 326 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_01.json.gz"
|
| 327 |
+
}
|
04_robocasa_scaling/scaling_n3_d1000_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n3_d1000_seed0",
|
| 3 |
+
"original_run_name": "e1_grid_d1000_n3_nodyn_seed0_100k_cuda0-4-5-7",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 1000,
|
| 7 |
+
"N": 3,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 100000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n3_d100_seed0/config.json
ADDED
|
@@ -0,0 +1,327 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 256,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 1024,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 8,
|
| 13 |
+
"n_layers": 4,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/E1_scaling_grid/configs_world2/e1_grid_d100_n3_nodyn_seed0_10k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 1024,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 4,
|
| 107 |
+
"n_heads": 8,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.0003,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 500,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "wl23_mg100_seed0",
|
| 125 |
+
"grad_accum_steps": 2,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 10000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 1024
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${EXPERIMENT_ROOT}/E1_scaling_grid/runs/e1_grid_d100_n3_nodyn_seed0_10k_cuda2-3",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.0003,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 1000,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_layers": 4,
|
| 270 |
+
"scaling_action_head_width": 256,
|
| 271 |
+
"scaling_analysis_suites": [
|
| 272 |
+
"e1_grid"
|
| 273 |
+
],
|
| 274 |
+
"scaling_backbone_layers": 8,
|
| 275 |
+
"scaling_encoder_layers": 4,
|
| 276 |
+
"scaling_expected_world_size": 2,
|
| 277 |
+
"scaling_family_width": 1024,
|
| 278 |
+
"scaling_global_batch_size": 192,
|
| 279 |
+
"scaling_grid_data_demos_per_task": 100,
|
| 280 |
+
"scaling_grid_model_size": "n3",
|
| 281 |
+
"scaling_grid_protocol": "direct_5x5_seed0_first_pass_v20260711",
|
| 282 |
+
"scaling_lr_action_head": 0.0003,
|
| 283 |
+
"scaling_lr_encoder": 0.0003,
|
| 284 |
+
"scaling_lr_lookup_model": "n3",
|
| 285 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 286 |
+
"scaling_lr_param_groups": [
|
| 287 |
+
"encoder",
|
| 288 |
+
"predictor",
|
| 289 |
+
"action_head"
|
| 290 |
+
],
|
| 291 |
+
"scaling_lr_predictor": 0.0003,
|
| 292 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 293 |
+
"scaling_lr_table_version": "n_lookup_v20260711",
|
| 294 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 295 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 296 |
+
"scaling_per_rank_microbatch": 48,
|
| 297 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 298 |
+
"scaling_recipe": "nodyn",
|
| 299 |
+
"scaling_run_name": "e1_grid_d100_n3_nodyn_seed0_10k",
|
| 300 |
+
"scaling_secondary_holdout_metric": "action_rmse",
|
| 301 |
+
"scaling_suite": "e1_grid",
|
| 302 |
+
"seed": 0,
|
| 303 |
+
"seq_len": 10,
|
| 304 |
+
"sequence_model": {
|
| 305 |
+
"backbone_type": "qwen2",
|
| 306 |
+
"hidden_dim": 1024,
|
| 307 |
+
"params": {
|
| 308 |
+
"attn_impl": "sdpa",
|
| 309 |
+
"dropout": 0.1,
|
| 310 |
+
"ffn_hidden_size": 2816,
|
| 311 |
+
"input_norm": false,
|
| 312 |
+
"max_context_len": 1024,
|
| 313 |
+
"model_dtype": "float32",
|
| 314 |
+
"n_heads": 8,
|
| 315 |
+
"n_kv_heads": null,
|
| 316 |
+
"n_layers": 8
|
| 317 |
+
},
|
| 318 |
+
"type": "continuous_transformer"
|
| 319 |
+
},
|
| 320 |
+
"use_proprio": true,
|
| 321 |
+
"warmup_steps": 500,
|
| 322 |
+
"weight_decay": 0.0,
|
| 323 |
+
"z_bottleneck": {
|
| 324 |
+
"dim": null
|
| 325 |
+
},
|
| 326 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_01.json.gz"
|
| 327 |
+
}
|
04_robocasa_scaling/scaling_n3_d100_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n3_d100_seed0",
|
| 3 |
+
"original_run_name": "e1_grid_d100_n3_nodyn_seed0_10k_cuda2-3",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 100,
|
| 7 |
+
"N": 3,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 10000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n3_d2900_seed0/config.json
ADDED
|
@@ -0,0 +1,327 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 256,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 1024,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 8,
|
| 13 |
+
"n_layers": 4,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/E1_scaling_grid/configs_world4/e1_grid_d2900_n3_nodyn_seed0_280k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 1024,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 4,
|
| 107 |
+
"n_heads": 8,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.0003,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 14000,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "wl23_mg2900_seed0",
|
| 125 |
+
"grad_accum_steps": 1,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 280000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 1024
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${EXPERIMENT_ROOT}/E1_scaling_grid/runs/e1_grid_d2900_n3_nodyn_seed0_280k_cuda1-2-3-6",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.0003,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 28000,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_layers": 4,
|
| 270 |
+
"scaling_action_head_width": 256,
|
| 271 |
+
"scaling_analysis_suites": [
|
| 272 |
+
"e1_grid"
|
| 273 |
+
],
|
| 274 |
+
"scaling_backbone_layers": 8,
|
| 275 |
+
"scaling_encoder_layers": 4,
|
| 276 |
+
"scaling_expected_world_size": 4,
|
| 277 |
+
"scaling_family_width": 1024,
|
| 278 |
+
"scaling_global_batch_size": 192,
|
| 279 |
+
"scaling_grid_data_demos_per_task": 2900,
|
| 280 |
+
"scaling_grid_model_size": "n3",
|
| 281 |
+
"scaling_grid_protocol": "direct_5x5_seed0_first_pass_v20260711",
|
| 282 |
+
"scaling_lr_action_head": 0.0003,
|
| 283 |
+
"scaling_lr_encoder": 0.0003,
|
| 284 |
+
"scaling_lr_lookup_model": "n3",
|
| 285 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 286 |
+
"scaling_lr_param_groups": [
|
| 287 |
+
"encoder",
|
| 288 |
+
"predictor",
|
| 289 |
+
"action_head"
|
| 290 |
+
],
|
| 291 |
+
"scaling_lr_predictor": 0.0003,
|
| 292 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 293 |
+
"scaling_lr_table_version": "n_lookup_v20260711",
|
| 294 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 295 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 296 |
+
"scaling_per_rank_microbatch": 48,
|
| 297 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 298 |
+
"scaling_recipe": "nodyn",
|
| 299 |
+
"scaling_run_name": "e1_grid_d2900_n3_nodyn_seed0_280k",
|
| 300 |
+
"scaling_secondary_holdout_metric": "action_rmse",
|
| 301 |
+
"scaling_suite": "e1_grid",
|
| 302 |
+
"seed": 0,
|
| 303 |
+
"seq_len": 10,
|
| 304 |
+
"sequence_model": {
|
| 305 |
+
"backbone_type": "qwen2",
|
| 306 |
+
"hidden_dim": 1024,
|
| 307 |
+
"params": {
|
| 308 |
+
"attn_impl": "sdpa",
|
| 309 |
+
"dropout": 0.1,
|
| 310 |
+
"ffn_hidden_size": 2816,
|
| 311 |
+
"input_norm": false,
|
| 312 |
+
"max_context_len": 1024,
|
| 313 |
+
"model_dtype": "float32",
|
| 314 |
+
"n_heads": 8,
|
| 315 |
+
"n_kv_heads": null,
|
| 316 |
+
"n_layers": 8
|
| 317 |
+
},
|
| 318 |
+
"type": "continuous_transformer"
|
| 319 |
+
},
|
| 320 |
+
"use_proprio": true,
|
| 321 |
+
"warmup_steps": 14000,
|
| 322 |
+
"weight_decay": 0.0,
|
| 323 |
+
"z_bottleneck": {
|
| 324 |
+
"dim": null
|
| 325 |
+
},
|
| 326 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_01.json.gz"
|
| 327 |
+
}
|
04_robocasa_scaling/scaling_n3_d2900_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n3_d2900_seed0",
|
| 3 |
+
"original_run_name": "e1_grid_d2900_n3_nodyn_seed0_280k_cuda1-2-3-6",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 2900,
|
| 7 |
+
"N": 3,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 280000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n3_d300_seed0/config.json
ADDED
|
@@ -0,0 +1,328 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 256,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 1024,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 8,
|
| 13 |
+
"n_layers": 4,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/E1_scaling_grid/configs_world1/e1_grid_d300_n3_nodyn_seed0_30k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 1024,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 4,
|
| 107 |
+
"n_heads": 8,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.0003,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 1500,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "300_demos",
|
| 125 |
+
"grad_accum_steps": 4,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 30000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 1024
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${EXPERIMENT_ROOT}/E1_scaling_grid/runs/e1_grid_d300_n3_nodyn_seed0_30k_cuda0",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.0003,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 3000,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_layers": 4,
|
| 270 |
+
"scaling_action_head_width": 256,
|
| 271 |
+
"scaling_analysis_suites": [
|
| 272 |
+
"e1_grid"
|
| 273 |
+
],
|
| 274 |
+
"scaling_backbone_layers": 8,
|
| 275 |
+
"scaling_encoder_layers": 4,
|
| 276 |
+
"scaling_expected_world_size": 1,
|
| 277 |
+
"scaling_family_width": 1024,
|
| 278 |
+
"scaling_global_batch_size": 192,
|
| 279 |
+
"scaling_grid_data_demos_per_task": 300,
|
| 280 |
+
"scaling_grid_model_size": "n3",
|
| 281 |
+
"scaling_grid_protocol": "direct_5x5_seed0_first_pass_v20260711",
|
| 282 |
+
"scaling_lr_action_head": 0.0003,
|
| 283 |
+
"scaling_lr_encoder": 0.0003,
|
| 284 |
+
"scaling_lr_lookup_model": "n3",
|
| 285 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 286 |
+
"scaling_lr_param_groups": [
|
| 287 |
+
"encoder",
|
| 288 |
+
"predictor",
|
| 289 |
+
"action_head"
|
| 290 |
+
],
|
| 291 |
+
"scaling_lr_predictor": 0.0003,
|
| 292 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 293 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 294 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 295 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 296 |
+
"scaling_model_size": "n3",
|
| 297 |
+
"scaling_per_rank_microbatch": 48,
|
| 298 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 299 |
+
"scaling_recipe": "nodyn",
|
| 300 |
+
"scaling_run_name": "e1_grid_d300_n3_nodyn_seed0_30k",
|
| 301 |
+
"scaling_secondary_holdout_metric": "action_rmse",
|
| 302 |
+
"scaling_suite": "e1_grid",
|
| 303 |
+
"seed": 0,
|
| 304 |
+
"seq_len": 10,
|
| 305 |
+
"sequence_model": {
|
| 306 |
+
"backbone_type": "qwen2",
|
| 307 |
+
"hidden_dim": 1024,
|
| 308 |
+
"params": {
|
| 309 |
+
"attn_impl": "sdpa",
|
| 310 |
+
"dropout": 0.1,
|
| 311 |
+
"ffn_hidden_size": 2816,
|
| 312 |
+
"input_norm": false,
|
| 313 |
+
"max_context_len": 1024,
|
| 314 |
+
"model_dtype": "float32",
|
| 315 |
+
"n_heads": 8,
|
| 316 |
+
"n_kv_heads": null,
|
| 317 |
+
"n_layers": 8
|
| 318 |
+
},
|
| 319 |
+
"type": "continuous_transformer"
|
| 320 |
+
},
|
| 321 |
+
"use_proprio": true,
|
| 322 |
+
"warmup_steps": 1500,
|
| 323 |
+
"weight_decay": 0.0,
|
| 324 |
+
"z_bottleneck": {
|
| 325 |
+
"dim": null
|
| 326 |
+
},
|
| 327 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_01.json.gz"
|
| 328 |
+
}
|
04_robocasa_scaling/scaling_n3_d300_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n3_d300_seed0",
|
| 3 |
+
"original_run_name": "e1_grid_d300_n3_nodyn_seed0_30k_cuda0",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 300,
|
| 7 |
+
"N": 3,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 30000
|
| 13 |
+
}
|
04_robocasa_scaling/scaling_n3_d50_seed0/config.json
ADDED
|
@@ -0,0 +1,327 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 256,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 1024,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 8,
|
| 13 |
+
"n_layers": 4,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/E1_scaling_grid/configs_world1/e1_grid_d50_n3_nodyn_seed0_5k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_data": false,
|
| 85 |
+
"dynamics": {
|
| 86 |
+
"enabled": false,
|
| 87 |
+
"params": {},
|
| 88 |
+
"type": "none"
|
| 89 |
+
},
|
| 90 |
+
"dynamics_decoder_type": "film",
|
| 91 |
+
"enable_pred_next": false,
|
| 92 |
+
"encoder": {
|
| 93 |
+
"params": {
|
| 94 |
+
"cnn_depth": 48,
|
| 95 |
+
"cnn_kernel": 5,
|
| 96 |
+
"cnn_mults": [
|
| 97 |
+
2,
|
| 98 |
+
3,
|
| 99 |
+
4,
|
| 100 |
+
8,
|
| 101 |
+
16
|
| 102 |
+
],
|
| 103 |
+
"d_model": 1024,
|
| 104 |
+
"dropout": 0.0,
|
| 105 |
+
"mlp_ratio": 4,
|
| 106 |
+
"n_fusion_layers": 4,
|
| 107 |
+
"n_heads": 8,
|
| 108 |
+
"readout_depth": 1,
|
| 109 |
+
"readout_mlp_ratio": 2,
|
| 110 |
+
"readout_queries": 4,
|
| 111 |
+
"rope_theta": 10000.0,
|
| 112 |
+
"use_proprio": true
|
| 113 |
+
},
|
| 114 |
+
"type": "attn_fusion_latent_token"
|
| 115 |
+
},
|
| 116 |
+
"encoder_lr": 0.0003,
|
| 117 |
+
"eval_at_start": false,
|
| 118 |
+
"eval_batch_size": 64,
|
| 119 |
+
"eval_crops_per_demo": 8,
|
| 120 |
+
"eval_every": 250,
|
| 121 |
+
"eval_num_workers": 24,
|
| 122 |
+
"eval_seed": 0,
|
| 123 |
+
"ffn_hidden_size": 1024,
|
| 124 |
+
"filter_key": "wl23_mg50_seed0",
|
| 125 |
+
"grad_accum_steps": 4,
|
| 126 |
+
"grad_clip": 1.0,
|
| 127 |
+
"hdf5_paths": [
|
| 128 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 129 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 130 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 151 |
+
],
|
| 152 |
+
"holdout_action_mean_samples": 1,
|
| 153 |
+
"holdout_crops_per_demo": 8,
|
| 154 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 155 |
+
"holdout_per_task_metrics": true,
|
| 156 |
+
"holdout_rmse_samplers": [
|
| 157 |
+
"deterministic",
|
| 158 |
+
"stochastic"
|
| 159 |
+
],
|
| 160 |
+
"holdout_sample_deterministic": null,
|
| 161 |
+
"holdout_sample_prefix_horizon": 4,
|
| 162 |
+
"holdout_sample_rmse": true,
|
| 163 |
+
"holdout_size": 2300,
|
| 164 |
+
"image_emb_dim": 512,
|
| 165 |
+
"init_device": "meta",
|
| 166 |
+
"input_norm": true,
|
| 167 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 168 |
+
"lang_emb_mode": "clip",
|
| 169 |
+
"lang_encoder_device": "auto",
|
| 170 |
+
"lang_obs_emb_dim": 256,
|
| 171 |
+
"log_every": 25,
|
| 172 |
+
"lora": false,
|
| 173 |
+
"lora_alpha": 32,
|
| 174 |
+
"lora_dropout": 0.0,
|
| 175 |
+
"lora_rank": 16,
|
| 176 |
+
"lora_target": [
|
| 177 |
+
"q_proj",
|
| 178 |
+
"k_proj",
|
| 179 |
+
"v_proj",
|
| 180 |
+
"o_proj",
|
| 181 |
+
"gate_proj",
|
| 182 |
+
"up_proj",
|
| 183 |
+
"down_proj"
|
| 184 |
+
],
|
| 185 |
+
"lr": 0.0003,
|
| 186 |
+
"max_context_len": 1024,
|
| 187 |
+
"max_steps": 5000,
|
| 188 |
+
"min_lr_ratio": 0.1,
|
| 189 |
+
"model": {
|
| 190 |
+
"action_chunk_len": 10,
|
| 191 |
+
"latent_dim": 1024
|
| 192 |
+
},
|
| 193 |
+
"model_dtype": "float32",
|
| 194 |
+
"n_heads": 16,
|
| 195 |
+
"n_kv_heads": null,
|
| 196 |
+
"n_layers": 2,
|
| 197 |
+
"norm_bias_weight_decay": null,
|
| 198 |
+
"norm_fit_batches": 0,
|
| 199 |
+
"num_workers": 24,
|
| 200 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 201 |
+
"obs_spec": {
|
| 202 |
+
"image_channels": 3,
|
| 203 |
+
"image_hw": [
|
| 204 |
+
128,
|
| 205 |
+
128
|
| 206 |
+
],
|
| 207 |
+
"image_keys": [
|
| 208 |
+
"robot0_agentview_left_image",
|
| 209 |
+
"robot0_agentview_right_image",
|
| 210 |
+
"robot0_eye_in_hand_image"
|
| 211 |
+
],
|
| 212 |
+
"lang_dim": 768,
|
| 213 |
+
"layout": "NHWC",
|
| 214 |
+
"low_dim_dims": [
|
| 215 |
+
3,
|
| 216 |
+
4,
|
| 217 |
+
3,
|
| 218 |
+
4,
|
| 219 |
+
2
|
| 220 |
+
],
|
| 221 |
+
"low_dim_keys": [
|
| 222 |
+
"robot0_base_to_eef_pos",
|
| 223 |
+
"robot0_base_to_eef_quat",
|
| 224 |
+
"robot0_base_pos",
|
| 225 |
+
"robot0_base_quat",
|
| 226 |
+
"robot0_gripper_qpos"
|
| 227 |
+
]
|
| 228 |
+
},
|
| 229 |
+
"obs_stride": 4,
|
| 230 |
+
"output_dir": "${EXPERIMENT_ROOT}/E1_scaling_grid/runs/e1_grid_d50_n3_nodyn_seed0_5k_cuda6",
|
| 231 |
+
"persistent_workers": true,
|
| 232 |
+
"precision": "bf16",
|
| 233 |
+
"pred_next_image_key_weights": {},
|
| 234 |
+
"pred_next_image_weight": 1.0,
|
| 235 |
+
"pred_next_lang_weight": 0.1,
|
| 236 |
+
"pred_next_mode": "terminal",
|
| 237 |
+
"pred_next_obs_offset": 1,
|
| 238 |
+
"pred_next_proprio_weight": 0.15,
|
| 239 |
+
"pred_next_steps": 4,
|
| 240 |
+
"pred_warmup_steps": 0,
|
| 241 |
+
"predictor_lr": 0.0003,
|
| 242 |
+
"prefetch_factor": 4,
|
| 243 |
+
"prepare_lang_cache_only": false,
|
| 244 |
+
"proprio_emb_dim": 128,
|
| 245 |
+
"recon": {
|
| 246 |
+
"enabled": false,
|
| 247 |
+
"params": {
|
| 248 |
+
"depth": 48,
|
| 249 |
+
"kernel_size": 5,
|
| 250 |
+
"mults": [
|
| 251 |
+
2,
|
| 252 |
+
3,
|
| 253 |
+
4,
|
| 254 |
+
4
|
| 255 |
+
]
|
| 256 |
+
}
|
| 257 |
+
},
|
| 258 |
+
"recon_image_weight": 1.0,
|
| 259 |
+
"recon_lang_weight": 0.1,
|
| 260 |
+
"recon_obs": false,
|
| 261 |
+
"recon_proprio_weight": 1.0,
|
| 262 |
+
"resume": null,
|
| 263 |
+
"resume_final_lr": null,
|
| 264 |
+
"resume_lr_schedule": "checkpoint",
|
| 265 |
+
"robomimic_src": null,
|
| 266 |
+
"save_every": 500,
|
| 267 |
+
"save_holdout_traces": true,
|
| 268 |
+
"save_latest_checkpoint": true,
|
| 269 |
+
"scaling_action_head_layers": 4,
|
| 270 |
+
"scaling_action_head_width": 256,
|
| 271 |
+
"scaling_analysis_suites": [
|
| 272 |
+
"e1_grid"
|
| 273 |
+
],
|
| 274 |
+
"scaling_backbone_layers": 8,
|
| 275 |
+
"scaling_encoder_layers": 4,
|
| 276 |
+
"scaling_expected_world_size": 1,
|
| 277 |
+
"scaling_family_width": 1024,
|
| 278 |
+
"scaling_global_batch_size": 192,
|
| 279 |
+
"scaling_grid_data_demos_per_task": 50,
|
| 280 |
+
"scaling_grid_model_size": "n3",
|
| 281 |
+
"scaling_grid_protocol": "direct_5x5_seed0_first_pass_v20260711",
|
| 282 |
+
"scaling_lr_action_head": 0.0003,
|
| 283 |
+
"scaling_lr_encoder": 0.0003,
|
| 284 |
+
"scaling_lr_lookup_model": "n3",
|
| 285 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 286 |
+
"scaling_lr_param_groups": [
|
| 287 |
+
"encoder",
|
| 288 |
+
"predictor",
|
| 289 |
+
"action_head"
|
| 290 |
+
],
|
| 291 |
+
"scaling_lr_predictor": 0.0003,
|
| 292 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 293 |
+
"scaling_lr_table_version": "n_lookup_v20260711",
|
| 294 |
+
"scaling_model_axis": "full_model_excluding_cnn_stem",
|
| 295 |
+
"scaling_model_family": "equiscale_encoder_backbone_head_v20260708",
|
| 296 |
+
"scaling_per_rank_microbatch": 48,
|
| 297 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 298 |
+
"scaling_recipe": "nodyn",
|
| 299 |
+
"scaling_run_name": "e1_grid_d50_n3_nodyn_seed0_5k",
|
| 300 |
+
"scaling_secondary_holdout_metric": "action_rmse",
|
| 301 |
+
"scaling_suite": "e1_grid",
|
| 302 |
+
"seed": 0,
|
| 303 |
+
"seq_len": 10,
|
| 304 |
+
"sequence_model": {
|
| 305 |
+
"backbone_type": "qwen2",
|
| 306 |
+
"hidden_dim": 1024,
|
| 307 |
+
"params": {
|
| 308 |
+
"attn_impl": "sdpa",
|
| 309 |
+
"dropout": 0.1,
|
| 310 |
+
"ffn_hidden_size": 2816,
|
| 311 |
+
"input_norm": false,
|
| 312 |
+
"max_context_len": 1024,
|
| 313 |
+
"model_dtype": "float32",
|
| 314 |
+
"n_heads": 8,
|
| 315 |
+
"n_kv_heads": null,
|
| 316 |
+
"n_layers": 8
|
| 317 |
+
},
|
| 318 |
+
"type": "continuous_transformer"
|
| 319 |
+
},
|
| 320 |
+
"use_proprio": true,
|
| 321 |
+
"warmup_steps": 250,
|
| 322 |
+
"weight_decay": 0.0,
|
| 323 |
+
"z_bottleneck": {
|
| 324 |
+
"dim": null
|
| 325 |
+
},
|
| 326 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_01.json.gz"
|
| 327 |
+
}
|
04_robocasa_scaling/scaling_n3_d50_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "scaling_n3_d50_seed0",
|
| 3 |
+
"original_run_name": "e1_grid_d50_n3_nodyn_seed0_5k_cuda6",
|
| 4 |
+
"paper_section": 4,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 50,
|
| 7 |
+
"N": 3,
|
| 8 |
+
"K": 1,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 5000
|
| 13 |
+
}
|
05_token_interface/README.md
ADDED
|
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Section 5 — Token interface
|
| 2 |
+
|
| 3 |
+
Paper: **Do more tokens per timestep improve performance?**
|
| 4 |
+
|
| 5 |
+
Ten final checkpoints: N2, K∈{4,50}, seed 0, C_train=10, across all five data settings.
|
| 6 |
+
|
| 7 |
+
| Run | Training steps | Checkpoint file |
|
| 8 |
+
|---|---:|---|
|
| 9 |
+
| [multitoken_n2_k4_d50_seed0](multitoken_n2_k4_d50_seed0/) | 5,000 | `checkpoint_step_00005000.pt` |
|
| 10 |
+
| [rawtoken_n2_k50_d50_seed0](rawtoken_n2_k50_d50_seed0/) | 5,000 | `checkpoint_step_00005000.pt` |
|
| 11 |
+
| [multitoken_n2_k4_d100_seed0](multitoken_n2_k4_d100_seed0/) | 10,000 | `checkpoint_step_00010000.pt` |
|
| 12 |
+
| [rawtoken_n2_k50_d100_seed0](rawtoken_n2_k50_d100_seed0/) | 10,000 | `checkpoint_step_00010000.pt` |
|
| 13 |
+
| [multitoken_n2_k4_d300_seed0](multitoken_n2_k4_d300_seed0/) | 30,000 | `checkpoint_step_00030000.pt` |
|
| 14 |
+
| [rawtoken_n2_k50_d300_seed0](rawtoken_n2_k50_d300_seed0/) | 30,000 | `checkpoint_step_00030000.pt` |
|
| 15 |
+
| [multitoken_n2_k4_d1000_seed0](multitoken_n2_k4_d1000_seed0/) | 100,000 | `checkpoint_step_00100000.pt` |
|
| 16 |
+
| [rawtoken_n2_k50_d1000_seed0](rawtoken_n2_k50_d1000_seed0/) | 100,000 | `checkpoint_step_00100000.pt` |
|
| 17 |
+
| [multitoken_n2_k4_d2900_seed0](multitoken_n2_k4_d2900_seed0/) | 280,000 | `checkpoint_step_00280000.pt` |
|
| 18 |
+
| [rawtoken_n2_k50_d2900_seed0](rawtoken_n2_k50_d2900_seed0/) | 280,000 | `checkpoint_step_00280000.pt` |
|
| 19 |
+
|
| 20 |
+
The K=1 controls reuse the N2 / seed-0 checkpoints in [Section 4](../04_robocasa_scaling/README.md).
|
05_token_interface/multitoken_n2_k4_d1000_seed0/config.json
ADDED
|
@@ -0,0 +1,337 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 192,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 768,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 6,
|
| 13 |
+
"n_layers": 2,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/E_multitoken_n2/configs_world4/multitoken_d1000_n2_k4_nodyn_seed0_100k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_batch_size": 1,
|
| 85 |
+
"dry_run_data": false,
|
| 86 |
+
"dynamics": {
|
| 87 |
+
"enabled": false,
|
| 88 |
+
"params": {},
|
| 89 |
+
"type": "none"
|
| 90 |
+
},
|
| 91 |
+
"dynamics_decoder_type": "film",
|
| 92 |
+
"enable_pred_next": false,
|
| 93 |
+
"encoder": {
|
| 94 |
+
"params": {
|
| 95 |
+
"cnn_depth": 48,
|
| 96 |
+
"cnn_kernel": 5,
|
| 97 |
+
"cnn_mults": [
|
| 98 |
+
2,
|
| 99 |
+
3,
|
| 100 |
+
4,
|
| 101 |
+
8,
|
| 102 |
+
16
|
| 103 |
+
],
|
| 104 |
+
"d_model": 768,
|
| 105 |
+
"dropout": 0.0,
|
| 106 |
+
"latent_token_self_attn": "full",
|
| 107 |
+
"mlp_ratio": 4,
|
| 108 |
+
"n_fusion_layers": 2,
|
| 109 |
+
"n_heads": 6,
|
| 110 |
+
"readout_depth": 1,
|
| 111 |
+
"readout_mlp_ratio": 2,
|
| 112 |
+
"readout_queries": 4,
|
| 113 |
+
"retained_tokens_per_frame": 4,
|
| 114 |
+
"rope_theta": 10000.0,
|
| 115 |
+
"use_proprio": true
|
| 116 |
+
},
|
| 117 |
+
"type": "attn_fusion_multi_token"
|
| 118 |
+
},
|
| 119 |
+
"encoder_lr": 0.000425,
|
| 120 |
+
"eval_at_start": false,
|
| 121 |
+
"eval_batch_size": 64,
|
| 122 |
+
"eval_crops_per_demo": 8,
|
| 123 |
+
"eval_every": 5000,
|
| 124 |
+
"eval_num_workers": 24,
|
| 125 |
+
"eval_seed": 0,
|
| 126 |
+
"ffn_hidden_size": 1024,
|
| 127 |
+
"filter_key": "wl23_mg1000_seed0",
|
| 128 |
+
"grad_accum_steps": 1,
|
| 129 |
+
"grad_clip": 1.0,
|
| 130 |
+
"hdf5_paths": [
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 151 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 152 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 153 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 154 |
+
],
|
| 155 |
+
"holdout_action_mean_samples": 1,
|
| 156 |
+
"holdout_crops_per_demo": 8,
|
| 157 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 158 |
+
"holdout_per_task_metrics": true,
|
| 159 |
+
"holdout_rmse_samplers": [
|
| 160 |
+
"deterministic",
|
| 161 |
+
"stochastic"
|
| 162 |
+
],
|
| 163 |
+
"holdout_sample_deterministic": null,
|
| 164 |
+
"holdout_sample_prefix_horizon": 4,
|
| 165 |
+
"holdout_sample_rmse": true,
|
| 166 |
+
"holdout_size": 2300,
|
| 167 |
+
"image_emb_dim": 512,
|
| 168 |
+
"init_device": "meta",
|
| 169 |
+
"input_norm": true,
|
| 170 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 171 |
+
"lang_emb_mode": "clip",
|
| 172 |
+
"lang_encoder_device": "auto",
|
| 173 |
+
"lang_obs_emb_dim": 256,
|
| 174 |
+
"log_every": 25,
|
| 175 |
+
"lora": false,
|
| 176 |
+
"lora_alpha": 32,
|
| 177 |
+
"lora_dropout": 0.0,
|
| 178 |
+
"lora_rank": 16,
|
| 179 |
+
"lora_target": [
|
| 180 |
+
"q_proj",
|
| 181 |
+
"k_proj",
|
| 182 |
+
"v_proj",
|
| 183 |
+
"o_proj",
|
| 184 |
+
"gate_proj",
|
| 185 |
+
"up_proj",
|
| 186 |
+
"down_proj"
|
| 187 |
+
],
|
| 188 |
+
"lr": 0.0003,
|
| 189 |
+
"max_context_len": 1024,
|
| 190 |
+
"max_steps": 100000,
|
| 191 |
+
"min_lr_ratio": 0.1,
|
| 192 |
+
"model": {
|
| 193 |
+
"action_chunk_len": 10,
|
| 194 |
+
"latent_dim": 768
|
| 195 |
+
},
|
| 196 |
+
"model_dtype": "float32",
|
| 197 |
+
"n_heads": 16,
|
| 198 |
+
"n_kv_heads": null,
|
| 199 |
+
"n_layers": 2,
|
| 200 |
+
"norm_bias_weight_decay": null,
|
| 201 |
+
"norm_fit_batches": 0,
|
| 202 |
+
"num_workers": 24,
|
| 203 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 204 |
+
"obs_spec": {
|
| 205 |
+
"image_channels": 3,
|
| 206 |
+
"image_hw": [
|
| 207 |
+
128,
|
| 208 |
+
128
|
| 209 |
+
],
|
| 210 |
+
"image_keys": [
|
| 211 |
+
"robot0_agentview_left_image",
|
| 212 |
+
"robot0_agentview_right_image",
|
| 213 |
+
"robot0_eye_in_hand_image"
|
| 214 |
+
],
|
| 215 |
+
"lang_dim": 768,
|
| 216 |
+
"layout": "NHWC",
|
| 217 |
+
"low_dim_dims": [
|
| 218 |
+
3,
|
| 219 |
+
4,
|
| 220 |
+
3,
|
| 221 |
+
4,
|
| 222 |
+
2
|
| 223 |
+
],
|
| 224 |
+
"low_dim_keys": [
|
| 225 |
+
"robot0_base_to_eef_pos",
|
| 226 |
+
"robot0_base_to_eef_quat",
|
| 227 |
+
"robot0_base_pos",
|
| 228 |
+
"robot0_base_quat",
|
| 229 |
+
"robot0_gripper_qpos"
|
| 230 |
+
]
|
| 231 |
+
},
|
| 232 |
+
"obs_stride": 4,
|
| 233 |
+
"output_dir": "${RELEASE_ROOT}/05_token_interface/multitoken_n2_k4_d1000_seed0",
|
| 234 |
+
"persistent_workers": true,
|
| 235 |
+
"precision": "bf16",
|
| 236 |
+
"pred_next_image_key_weights": {},
|
| 237 |
+
"pred_next_image_weight": 1.0,
|
| 238 |
+
"pred_next_lang_weight": 0.1,
|
| 239 |
+
"pred_next_mode": "terminal",
|
| 240 |
+
"pred_next_obs_offset": 1,
|
| 241 |
+
"pred_next_proprio_weight": 0.15,
|
| 242 |
+
"pred_next_steps": 4,
|
| 243 |
+
"pred_warmup_steps": 0,
|
| 244 |
+
"predictor_lr": 0.000425,
|
| 245 |
+
"prefetch_factor": 4,
|
| 246 |
+
"prepare_lang_cache_only": false,
|
| 247 |
+
"proprio_emb_dim": 128,
|
| 248 |
+
"recon": {
|
| 249 |
+
"enabled": false,
|
| 250 |
+
"params": {
|
| 251 |
+
"depth": 48,
|
| 252 |
+
"kernel_size": 5,
|
| 253 |
+
"mults": [
|
| 254 |
+
2,
|
| 255 |
+
3,
|
| 256 |
+
4,
|
| 257 |
+
4
|
| 258 |
+
]
|
| 259 |
+
}
|
| 260 |
+
},
|
| 261 |
+
"recon_image_weight": 1.0,
|
| 262 |
+
"recon_lang_weight": 0.1,
|
| 263 |
+
"recon_obs": false,
|
| 264 |
+
"recon_proprio_weight": 1.0,
|
| 265 |
+
"resume": null,
|
| 266 |
+
"resume_final_lr": null,
|
| 267 |
+
"resume_lr_schedule": "checkpoint",
|
| 268 |
+
"robomimic_src": "${EXTERNAL_ROOT}/asset_977d654eb1/robomimic",
|
| 269 |
+
"save_every": 10000,
|
| 270 |
+
"save_holdout_traces": true,
|
| 271 |
+
"save_latest_checkpoint": true,
|
| 272 |
+
"scaling_action_head_d_model": 192,
|
| 273 |
+
"scaling_action_head_layers": 2,
|
| 274 |
+
"scaling_action_head_width": 192,
|
| 275 |
+
"scaling_analysis_suites": [
|
| 276 |
+
"world_tokens_per_frame_ablation"
|
| 277 |
+
],
|
| 278 |
+
"scaling_backbone_d_model": 768,
|
| 279 |
+
"scaling_backbone_layers": 4,
|
| 280 |
+
"scaling_encoder_d_model": 768,
|
| 281 |
+
"scaling_encoder_layers": 2,
|
| 282 |
+
"scaling_expected_world_size": 4,
|
| 283 |
+
"scaling_family_width": 768,
|
| 284 |
+
"scaling_global_batch_size": 192,
|
| 285 |
+
"scaling_grid_data_demos_per_task": 1000,
|
| 286 |
+
"scaling_grid_model_size": "n2",
|
| 287 |
+
"scaling_grid_protocol": "multitoken_d1000_n2_v1",
|
| 288 |
+
"scaling_latent_dim": 768,
|
| 289 |
+
"scaling_lr_action_head": 0.0003,
|
| 290 |
+
"scaling_lr_encoder": 0.000425,
|
| 291 |
+
"scaling_lr_lookup_model": "n2",
|
| 292 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 293 |
+
"scaling_lr_param_groups": [
|
| 294 |
+
"encoder",
|
| 295 |
+
"predictor",
|
| 296 |
+
"action_head"
|
| 297 |
+
],
|
| 298 |
+
"scaling_lr_predictor": 0.000425,
|
| 299 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 300 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 301 |
+
"scaling_model_axis": "retained_world_tokens_per_frame",
|
| 302 |
+
"scaling_model_family": "n2_multitoken_interface_v1",
|
| 303 |
+
"scaling_model_size": "n2",
|
| 304 |
+
"scaling_per_rank_microbatch": 48,
|
| 305 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 306 |
+
"scaling_recipe": "multitoken_causal_nodyn",
|
| 307 |
+
"scaling_run_name": "multitoken_d1000_n2_k4_nodyn_seed0_100k",
|
| 308 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/full10",
|
| 309 |
+
"scaling_suite": "world_tokens_per_frame_ablation",
|
| 310 |
+
"seed": 0,
|
| 311 |
+
"seq_len": 10,
|
| 312 |
+
"sequence_model": {
|
| 313 |
+
"backbone_type": "qwen2",
|
| 314 |
+
"hidden_dim": 768,
|
| 315 |
+
"params": {
|
| 316 |
+
"attn_impl": "sdpa",
|
| 317 |
+
"dropout": 0.1,
|
| 318 |
+
"ffn_hidden_size": 2048,
|
| 319 |
+
"frame_readout_index": -1,
|
| 320 |
+
"input_norm": false,
|
| 321 |
+
"max_context_len": 1024,
|
| 322 |
+
"model_dtype": "float32",
|
| 323 |
+
"n_heads": 6,
|
| 324 |
+
"n_kv_heads": null,
|
| 325 |
+
"n_layers": 4,
|
| 326 |
+
"tokens_per_frame": 4
|
| 327 |
+
},
|
| 328 |
+
"type": "frame_major_continuous_transformer"
|
| 329 |
+
},
|
| 330 |
+
"use_proprio": true,
|
| 331 |
+
"warmup_steps": 5000,
|
| 332 |
+
"weight_decay": 0.0,
|
| 333 |
+
"z_bottleneck": {
|
| 334 |
+
"dim": null
|
| 335 |
+
},
|
| 336 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_01.json.gz"
|
| 337 |
+
}
|
05_token_interface/multitoken_n2_k4_d1000_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "multitoken_n2_k4_d1000_seed0",
|
| 3 |
+
"original_run_name": "multitoken_d1000_n2_k4_nodyn_seed0_100k_cuda4-5-6-7",
|
| 4 |
+
"paper_section": 5,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 1000,
|
| 7 |
+
"N": 2,
|
| 8 |
+
"K": 4,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 100000
|
| 13 |
+
}
|
05_token_interface/multitoken_n2_k4_d100_seed0/config.json
ADDED
|
@@ -0,0 +1,333 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 192,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 768,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 6,
|
| 13 |
+
"n_layers": 2,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/E_multitoken_n2/configs_world1/multitoken_d100_n2_k4_nodyn_seed0_10k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_batch_size": 1,
|
| 85 |
+
"dry_run_data": false,
|
| 86 |
+
"dynamics": {
|
| 87 |
+
"enabled": false,
|
| 88 |
+
"params": {},
|
| 89 |
+
"type": "none"
|
| 90 |
+
},
|
| 91 |
+
"dynamics_decoder_type": "film",
|
| 92 |
+
"enable_pred_next": false,
|
| 93 |
+
"encoder": {
|
| 94 |
+
"params": {
|
| 95 |
+
"cnn_depth": 48,
|
| 96 |
+
"cnn_kernel": 5,
|
| 97 |
+
"cnn_mults": [
|
| 98 |
+
2,
|
| 99 |
+
3,
|
| 100 |
+
4,
|
| 101 |
+
8,
|
| 102 |
+
16
|
| 103 |
+
],
|
| 104 |
+
"d_model": 768,
|
| 105 |
+
"dropout": 0.0,
|
| 106 |
+
"latent_token_self_attn": "full",
|
| 107 |
+
"mlp_ratio": 4,
|
| 108 |
+
"n_fusion_layers": 2,
|
| 109 |
+
"n_heads": 6,
|
| 110 |
+
"readout_depth": 1,
|
| 111 |
+
"readout_mlp_ratio": 2,
|
| 112 |
+
"readout_queries": 4,
|
| 113 |
+
"retained_tokens_per_frame": 4,
|
| 114 |
+
"rope_theta": 10000.0,
|
| 115 |
+
"use_proprio": true
|
| 116 |
+
},
|
| 117 |
+
"type": "attn_fusion_multi_token"
|
| 118 |
+
},
|
| 119 |
+
"encoder_lr": 0.000425,
|
| 120 |
+
"eval_at_start": false,
|
| 121 |
+
"eval_batch_size": 64,
|
| 122 |
+
"eval_crops_per_demo": 8,
|
| 123 |
+
"eval_every": 500,
|
| 124 |
+
"eval_num_workers": 24,
|
| 125 |
+
"eval_seed": 0,
|
| 126 |
+
"ffn_hidden_size": 1024,
|
| 127 |
+
"filter_key": "wl23_mg100_seed0",
|
| 128 |
+
"grad_accum_steps": 4,
|
| 129 |
+
"grad_clip": 1.0,
|
| 130 |
+
"hdf5_paths": [
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 151 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 152 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 153 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 154 |
+
],
|
| 155 |
+
"holdout_action_mean_samples": 1,
|
| 156 |
+
"holdout_crops_per_demo": 8,
|
| 157 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 158 |
+
"holdout_per_task_metrics": true,
|
| 159 |
+
"holdout_rmse_samplers": [
|
| 160 |
+
"deterministic",
|
| 161 |
+
"stochastic"
|
| 162 |
+
],
|
| 163 |
+
"holdout_sample_deterministic": null,
|
| 164 |
+
"holdout_sample_prefix_horizon": 4,
|
| 165 |
+
"holdout_sample_rmse": true,
|
| 166 |
+
"holdout_size": 2300,
|
| 167 |
+
"image_emb_dim": 512,
|
| 168 |
+
"init_device": "meta",
|
| 169 |
+
"input_norm": true,
|
| 170 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 171 |
+
"lang_emb_mode": "clip",
|
| 172 |
+
"lang_encoder_device": "auto",
|
| 173 |
+
"lang_obs_emb_dim": 256,
|
| 174 |
+
"log_every": 25,
|
| 175 |
+
"lora": false,
|
| 176 |
+
"lora_alpha": 32,
|
| 177 |
+
"lora_dropout": 0.0,
|
| 178 |
+
"lora_rank": 16,
|
| 179 |
+
"lora_target": [
|
| 180 |
+
"q_proj",
|
| 181 |
+
"k_proj",
|
| 182 |
+
"v_proj",
|
| 183 |
+
"o_proj",
|
| 184 |
+
"gate_proj",
|
| 185 |
+
"up_proj",
|
| 186 |
+
"down_proj"
|
| 187 |
+
],
|
| 188 |
+
"lr": 0.0003,
|
| 189 |
+
"max_context_len": 1024,
|
| 190 |
+
"max_steps": 10000,
|
| 191 |
+
"min_lr_ratio": 0.1,
|
| 192 |
+
"model": {
|
| 193 |
+
"action_chunk_len": 10,
|
| 194 |
+
"latent_dim": 768
|
| 195 |
+
},
|
| 196 |
+
"model_dtype": "float32",
|
| 197 |
+
"n_heads": 16,
|
| 198 |
+
"n_kv_heads": null,
|
| 199 |
+
"n_layers": 2,
|
| 200 |
+
"norm_bias_weight_decay": null,
|
| 201 |
+
"norm_fit_batches": 0,
|
| 202 |
+
"num_workers": 24,
|
| 203 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 204 |
+
"obs_spec": {
|
| 205 |
+
"image_channels": 3,
|
| 206 |
+
"image_hw": [
|
| 207 |
+
128,
|
| 208 |
+
128
|
| 209 |
+
],
|
| 210 |
+
"image_keys": [
|
| 211 |
+
"robot0_agentview_left_image",
|
| 212 |
+
"robot0_agentview_right_image",
|
| 213 |
+
"robot0_eye_in_hand_image"
|
| 214 |
+
],
|
| 215 |
+
"lang_dim": 768,
|
| 216 |
+
"layout": "NHWC",
|
| 217 |
+
"low_dim_dims": [
|
| 218 |
+
3,
|
| 219 |
+
4,
|
| 220 |
+
3,
|
| 221 |
+
4,
|
| 222 |
+
2
|
| 223 |
+
],
|
| 224 |
+
"low_dim_keys": [
|
| 225 |
+
"robot0_base_to_eef_pos",
|
| 226 |
+
"robot0_base_to_eef_quat",
|
| 227 |
+
"robot0_base_pos",
|
| 228 |
+
"robot0_base_quat",
|
| 229 |
+
"robot0_gripper_qpos"
|
| 230 |
+
]
|
| 231 |
+
},
|
| 232 |
+
"obs_stride": 4,
|
| 233 |
+
"output_dir": "${RELEASE_ROOT}/05_token_interface/multitoken_n2_k4_d100_seed0",
|
| 234 |
+
"persistent_workers": true,
|
| 235 |
+
"precision": "bf16",
|
| 236 |
+
"pred_next_image_key_weights": {},
|
| 237 |
+
"pred_next_image_weight": 1.0,
|
| 238 |
+
"pred_next_lang_weight": 0.1,
|
| 239 |
+
"pred_next_mode": "terminal",
|
| 240 |
+
"pred_next_obs_offset": 1,
|
| 241 |
+
"pred_next_proprio_weight": 0.15,
|
| 242 |
+
"pred_next_steps": 4,
|
| 243 |
+
"pred_warmup_steps": 0,
|
| 244 |
+
"predictor_lr": 0.000425,
|
| 245 |
+
"prefetch_factor": 4,
|
| 246 |
+
"prepare_lang_cache_only": false,
|
| 247 |
+
"proprio_emb_dim": 128,
|
| 248 |
+
"recon": {
|
| 249 |
+
"enabled": false,
|
| 250 |
+
"params": {
|
| 251 |
+
"depth": 48,
|
| 252 |
+
"kernel_size": 5,
|
| 253 |
+
"mults": [
|
| 254 |
+
2,
|
| 255 |
+
3,
|
| 256 |
+
4,
|
| 257 |
+
4
|
| 258 |
+
]
|
| 259 |
+
}
|
| 260 |
+
},
|
| 261 |
+
"recon_image_weight": 1.0,
|
| 262 |
+
"recon_lang_weight": 0.1,
|
| 263 |
+
"recon_obs": false,
|
| 264 |
+
"recon_proprio_weight": 1.0,
|
| 265 |
+
"resume": null,
|
| 266 |
+
"resume_final_lr": null,
|
| 267 |
+
"resume_lr_schedule": "checkpoint",
|
| 268 |
+
"robomimic_src": "${EXTERNAL_ROOT}/asset_977d654eb1/robomimic",
|
| 269 |
+
"save_every": 1000,
|
| 270 |
+
"save_holdout_traces": true,
|
| 271 |
+
"save_latest_checkpoint": true,
|
| 272 |
+
"scaling_action_head_layers": 2,
|
| 273 |
+
"scaling_action_head_width": 192,
|
| 274 |
+
"scaling_analysis_suites": [
|
| 275 |
+
"world_tokens_per_frame_ablation"
|
| 276 |
+
],
|
| 277 |
+
"scaling_backbone_layers": 4,
|
| 278 |
+
"scaling_encoder_layers": 2,
|
| 279 |
+
"scaling_expected_world_size": 1,
|
| 280 |
+
"scaling_family_width": 768,
|
| 281 |
+
"scaling_global_batch_size": 192,
|
| 282 |
+
"scaling_grid_data_demos_per_task": 100,
|
| 283 |
+
"scaling_grid_model_size": "n2",
|
| 284 |
+
"scaling_grid_protocol": "multitoken_d100_n2_v1",
|
| 285 |
+
"scaling_lr_action_head": 0.0003,
|
| 286 |
+
"scaling_lr_encoder": 0.000425,
|
| 287 |
+
"scaling_lr_lookup_model": "n2",
|
| 288 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 289 |
+
"scaling_lr_param_groups": [
|
| 290 |
+
"encoder",
|
| 291 |
+
"predictor",
|
| 292 |
+
"action_head"
|
| 293 |
+
],
|
| 294 |
+
"scaling_lr_predictor": 0.000425,
|
| 295 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 296 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 297 |
+
"scaling_model_axis": "retained_world_tokens_per_frame",
|
| 298 |
+
"scaling_model_family": "n2_multitoken_interface_v1",
|
| 299 |
+
"scaling_model_size": "n2",
|
| 300 |
+
"scaling_per_rank_microbatch": 48,
|
| 301 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 302 |
+
"scaling_recipe": "multitoken_causal_nodyn",
|
| 303 |
+
"scaling_run_name": "multitoken_d100_n2_k4_nodyn_seed0_10k",
|
| 304 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/full10",
|
| 305 |
+
"scaling_suite": "world_tokens_per_frame_ablation",
|
| 306 |
+
"seed": 0,
|
| 307 |
+
"seq_len": 10,
|
| 308 |
+
"sequence_model": {
|
| 309 |
+
"backbone_type": "qwen2",
|
| 310 |
+
"hidden_dim": 768,
|
| 311 |
+
"params": {
|
| 312 |
+
"attn_impl": "sdpa",
|
| 313 |
+
"dropout": 0.1,
|
| 314 |
+
"ffn_hidden_size": 2048,
|
| 315 |
+
"frame_readout_index": -1,
|
| 316 |
+
"input_norm": false,
|
| 317 |
+
"max_context_len": 1024,
|
| 318 |
+
"model_dtype": "float32",
|
| 319 |
+
"n_heads": 6,
|
| 320 |
+
"n_kv_heads": null,
|
| 321 |
+
"n_layers": 4,
|
| 322 |
+
"tokens_per_frame": 4
|
| 323 |
+
},
|
| 324 |
+
"type": "frame_major_continuous_transformer"
|
| 325 |
+
},
|
| 326 |
+
"use_proprio": true,
|
| 327 |
+
"warmup_steps": 500,
|
| 328 |
+
"weight_decay": 0.0,
|
| 329 |
+
"z_bottleneck": {
|
| 330 |
+
"dim": null
|
| 331 |
+
},
|
| 332 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_01.json.gz"
|
| 333 |
+
}
|
05_token_interface/multitoken_n2_k4_d100_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "multitoken_n2_k4_d100_seed0",
|
| 3 |
+
"original_run_name": "multitoken_d100_n2_k4_nodyn_seed0_10k_cuda7",
|
| 4 |
+
"paper_section": 5,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 100,
|
| 7 |
+
"N": 2,
|
| 8 |
+
"K": 4,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 10000
|
| 13 |
+
}
|
05_token_interface/multitoken_n2_k4_d2900_seed0/config.json
ADDED
|
@@ -0,0 +1,333 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"action_chunk_len": 10,
|
| 3 |
+
"action_decoder_emb": 256,
|
| 4 |
+
"action_head": {
|
| 5 |
+
"params": {
|
| 6 |
+
"activation": "gelu",
|
| 7 |
+
"d_model": 192,
|
| 8 |
+
"denoising_steps": 20,
|
| 9 |
+
"dim_feedforward": 768,
|
| 10 |
+
"dropout": 0.0,
|
| 11 |
+
"h_adaln_bottleneck": false,
|
| 12 |
+
"n_heads": 6,
|
| 13 |
+
"n_layers": 2,
|
| 14 |
+
"use_h_cross_attn": false,
|
| 15 |
+
"use_obs_cross_attn": false
|
| 16 |
+
},
|
| 17 |
+
"type": "diffusion_dit"
|
| 18 |
+
},
|
| 19 |
+
"action_head_lr": 0.0003,
|
| 20 |
+
"action_spec": {
|
| 21 |
+
"dim": 12,
|
| 22 |
+
"discrete_dims": [
|
| 23 |
+
6,
|
| 24 |
+
11
|
| 25 |
+
]
|
| 26 |
+
},
|
| 27 |
+
"action_weight": 1.0,
|
| 28 |
+
"adamw_beta1": 0.9,
|
| 29 |
+
"adamw_beta2": 0.95,
|
| 30 |
+
"adamw_eps": 1e-08,
|
| 31 |
+
"allow_holdout_overlap": false,
|
| 32 |
+
"attn_backend": "auto",
|
| 33 |
+
"attn_impl": "eager",
|
| 34 |
+
"backbone_lr": null,
|
| 35 |
+
"backbone_type": "qwen2",
|
| 36 |
+
"batch_size": 48,
|
| 37 |
+
"cnn_depth": 48,
|
| 38 |
+
"cnn_kernel": 5,
|
| 39 |
+
"cnn_mults": [
|
| 40 |
+
2,
|
| 41 |
+
3,
|
| 42 |
+
4,
|
| 43 |
+
4
|
| 44 |
+
],
|
| 45 |
+
"config": "${EXPERIMENT_ROOT}/E_multitoken_n2/configs_world4/multitoken_d2900_n2_k4_nodyn_seed0_280k.yaml",
|
| 46 |
+
"crops_per_demo": 8,
|
| 47 |
+
"d_model": 2048,
|
| 48 |
+
"dataset": [
|
| 49 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 50 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 51 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 52 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 53 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 54 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 55 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 56 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 57 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 58 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 59 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 60 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 61 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 62 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 63 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 64 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 65 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 66 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 67 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 68 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 69 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 70 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5",
|
| 71 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5"
|
| 72 |
+
],
|
| 73 |
+
"ddp_timeout_minutes": 180.0,
|
| 74 |
+
"denoising_steps": 20,
|
| 75 |
+
"device": "cuda",
|
| 76 |
+
"diffusion_mlp_dims": [
|
| 77 |
+
1024,
|
| 78 |
+
1024,
|
| 79 |
+
1024
|
| 80 |
+
],
|
| 81 |
+
"diffusion_residual_style": true,
|
| 82 |
+
"diffusion_time_dim": 16,
|
| 83 |
+
"dropout": 0.1,
|
| 84 |
+
"dry_run_batch_size": 1,
|
| 85 |
+
"dry_run_data": false,
|
| 86 |
+
"dynamics": {
|
| 87 |
+
"enabled": false,
|
| 88 |
+
"params": {},
|
| 89 |
+
"type": "none"
|
| 90 |
+
},
|
| 91 |
+
"dynamics_decoder_type": "film",
|
| 92 |
+
"enable_pred_next": false,
|
| 93 |
+
"encoder": {
|
| 94 |
+
"params": {
|
| 95 |
+
"cnn_depth": 48,
|
| 96 |
+
"cnn_kernel": 5,
|
| 97 |
+
"cnn_mults": [
|
| 98 |
+
2,
|
| 99 |
+
3,
|
| 100 |
+
4,
|
| 101 |
+
8,
|
| 102 |
+
16
|
| 103 |
+
],
|
| 104 |
+
"d_model": 768,
|
| 105 |
+
"dropout": 0.0,
|
| 106 |
+
"latent_token_self_attn": "full",
|
| 107 |
+
"mlp_ratio": 4,
|
| 108 |
+
"n_fusion_layers": 2,
|
| 109 |
+
"n_heads": 6,
|
| 110 |
+
"readout_depth": 1,
|
| 111 |
+
"readout_mlp_ratio": 2,
|
| 112 |
+
"readout_queries": 4,
|
| 113 |
+
"retained_tokens_per_frame": 4,
|
| 114 |
+
"rope_theta": 10000.0,
|
| 115 |
+
"use_proprio": true
|
| 116 |
+
},
|
| 117 |
+
"type": "attn_fusion_multi_token"
|
| 118 |
+
},
|
| 119 |
+
"encoder_lr": 0.000425,
|
| 120 |
+
"eval_at_start": false,
|
| 121 |
+
"eval_batch_size": 64,
|
| 122 |
+
"eval_crops_per_demo": 8,
|
| 123 |
+
"eval_every": 14000,
|
| 124 |
+
"eval_num_workers": 24,
|
| 125 |
+
"eval_seed": 0,
|
| 126 |
+
"ffn_hidden_size": 1024,
|
| 127 |
+
"filter_key": "wl23_mg2900_seed0",
|
| 128 |
+
"grad_accum_steps": 1,
|
| 129 |
+
"grad_clip": 1.0,
|
| 130 |
+
"hdf5_paths": [
|
| 131 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeePressButton/mg/2024-05-04-22-21-32/demo_gentex_im128_randcams.hdf5",
|
| 132 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeServeMug/mg/2024-05-04-22-21-50/demo_gentex_im128_randcams.hdf5",
|
| 133 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_coffee/CoffeeSetupMug/mg/2024-05-04-22-22-13_and_2024-05-08-05-52-13/demo_gentex_im128_randcams.hdf5",
|
| 134 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseDoubleDoor/mg/2024-05-04-22-22-42_and_2024-05-08-06-02-36/demo_gentex_im128_randcams.hdf5",
|
| 135 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/CloseSingleDoor/mg/2024-05-04-22-34-56/demo_gentex_im128_randcams.hdf5",
|
| 136 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_doors/OpenSingleDoor/mg/2024-05-04-22-37-39/demo_gentex_im128_randcams.hdf5",
|
| 137 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/CloseDrawer/mg/2024-05-09-09-32-19/demo_gentex_im128_randcams.hdf5",
|
| 138 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_drawer/OpenDrawer/mg/2024-05-04-22-38-42/demo_gentex_im128_randcams.hdf5",
|
| 139 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOffMicrowave/mg/2024-05-04-22-39-23/demo_gentex_im128_randcams.hdf5",
|
| 140 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_microwave/TurnOnMicrowave/mg/2024-05-04-22-40-00/demo_gentex_im128_randcams.hdf5",
|
| 141 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCabToCounter/mg/2024-07-12-04-33-29/demo_gentex_im128_randcams.hdf5",
|
| 142 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToCab/mg/2024-05-04-22-12-27_and_2024-05-07-07-39-33/demo_gentex_im128_randcams.hdf5",
|
| 143 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToMicrowave/mg/2024-05-04-22-13-21_and_2024-05-07-07-41-17/demo_gentex_im128_randcams.hdf5",
|
| 144 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToSink/mg/2024-05-04-22-14-06_and_2024-05-07-07-40-17/demo_gentex_im128_randcams.hdf5",
|
| 145 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPCounterToStove/mg/2024-05-04-22-14-20/demo_gentex_im128_randcams.hdf5",
|
| 146 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPMicrowaveToCounter/mg/2024-05-04-22-14-26_and_2024-05-07-07-41-42/demo_gentex_im128_randcams.hdf5",
|
| 147 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPSinkToCounter/mg/2024-05-04-22-14-34_and_2024-05-07-07-40-21/demo_gentex_im128_randcams.hdf5",
|
| 148 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_pnp/PnPStoveToCounter/mg/2024-05-04-22-14-40/demo_gentex_im128_randcams.hdf5",
|
| 149 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOffSinkFaucet/mg/2024-05-04-22-17-26/demo_gentex_im128_randcams.hdf5",
|
| 150 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnOnSinkFaucet/mg/2024-05-04-22-17-46/demo_gentex_im128_randcams.hdf5",
|
| 151 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_sink/TurnSinkSpout/mg/2024-05-09-09-31-12/demo_gentex_im128_randcams.hdf5",
|
| 152 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOffStove/mg/2024-05-08-09-20-45/demo_gentex_im128_randcams.hdf5",
|
| 153 |
+
"${DATA_ROOT}/robocasa/mg_im/v0.1/single_stage/kitchen_stove/TurnOnStove/mg/2024-05-08-09-20-31/demo_gentex_im128_randcams.hdf5"
|
| 154 |
+
],
|
| 155 |
+
"holdout_action_mean_samples": 1,
|
| 156 |
+
"holdout_crops_per_demo": 8,
|
| 157 |
+
"holdout_filter_key": "wl23_holdout100pt_seed0",
|
| 158 |
+
"holdout_per_task_metrics": true,
|
| 159 |
+
"holdout_rmse_samplers": [
|
| 160 |
+
"deterministic",
|
| 161 |
+
"stochastic"
|
| 162 |
+
],
|
| 163 |
+
"holdout_sample_deterministic": null,
|
| 164 |
+
"holdout_sample_prefix_horizon": 4,
|
| 165 |
+
"holdout_sample_rmse": true,
|
| 166 |
+
"holdout_size": 2300,
|
| 167 |
+
"image_emb_dim": 512,
|
| 168 |
+
"init_device": "meta",
|
| 169 |
+
"input_norm": true,
|
| 170 |
+
"lang_emb_cache": "${EXPERIMENT_ROOT}/robocasa_mg23_clip_lang_emb_cache.npz",
|
| 171 |
+
"lang_emb_mode": "clip",
|
| 172 |
+
"lang_encoder_device": "auto",
|
| 173 |
+
"lang_obs_emb_dim": 256,
|
| 174 |
+
"log_every": 25,
|
| 175 |
+
"lora": false,
|
| 176 |
+
"lora_alpha": 32,
|
| 177 |
+
"lora_dropout": 0.0,
|
| 178 |
+
"lora_rank": 16,
|
| 179 |
+
"lora_target": [
|
| 180 |
+
"q_proj",
|
| 181 |
+
"k_proj",
|
| 182 |
+
"v_proj",
|
| 183 |
+
"o_proj",
|
| 184 |
+
"gate_proj",
|
| 185 |
+
"up_proj",
|
| 186 |
+
"down_proj"
|
| 187 |
+
],
|
| 188 |
+
"lr": 0.0003,
|
| 189 |
+
"max_context_len": 1024,
|
| 190 |
+
"max_steps": 280000,
|
| 191 |
+
"min_lr_ratio": 0.1,
|
| 192 |
+
"model": {
|
| 193 |
+
"action_chunk_len": 10,
|
| 194 |
+
"latent_dim": 768
|
| 195 |
+
},
|
| 196 |
+
"model_dtype": "float32",
|
| 197 |
+
"n_heads": 16,
|
| 198 |
+
"n_kv_heads": null,
|
| 199 |
+
"n_layers": 2,
|
| 200 |
+
"norm_bias_weight_decay": null,
|
| 201 |
+
"norm_fit_batches": 0,
|
| 202 |
+
"num_workers": 24,
|
| 203 |
+
"objective": "robocasa_lang_as_obs_image_state_diffusion_action",
|
| 204 |
+
"obs_spec": {
|
| 205 |
+
"image_channels": 3,
|
| 206 |
+
"image_hw": [
|
| 207 |
+
128,
|
| 208 |
+
128
|
| 209 |
+
],
|
| 210 |
+
"image_keys": [
|
| 211 |
+
"robot0_agentview_left_image",
|
| 212 |
+
"robot0_agentview_right_image",
|
| 213 |
+
"robot0_eye_in_hand_image"
|
| 214 |
+
],
|
| 215 |
+
"lang_dim": 768,
|
| 216 |
+
"layout": "NHWC",
|
| 217 |
+
"low_dim_dims": [
|
| 218 |
+
3,
|
| 219 |
+
4,
|
| 220 |
+
3,
|
| 221 |
+
4,
|
| 222 |
+
2
|
| 223 |
+
],
|
| 224 |
+
"low_dim_keys": [
|
| 225 |
+
"robot0_base_to_eef_pos",
|
| 226 |
+
"robot0_base_to_eef_quat",
|
| 227 |
+
"robot0_base_pos",
|
| 228 |
+
"robot0_base_quat",
|
| 229 |
+
"robot0_gripper_qpos"
|
| 230 |
+
]
|
| 231 |
+
},
|
| 232 |
+
"obs_stride": 4,
|
| 233 |
+
"output_dir": "${RELEASE_ROOT}/05_token_interface/multitoken_n2_k4_d2900_seed0",
|
| 234 |
+
"persistent_workers": true,
|
| 235 |
+
"precision": "bf16",
|
| 236 |
+
"pred_next_image_key_weights": {},
|
| 237 |
+
"pred_next_image_weight": 1.0,
|
| 238 |
+
"pred_next_lang_weight": 0.1,
|
| 239 |
+
"pred_next_mode": "terminal",
|
| 240 |
+
"pred_next_obs_offset": 1,
|
| 241 |
+
"pred_next_proprio_weight": 0.15,
|
| 242 |
+
"pred_next_steps": 4,
|
| 243 |
+
"pred_warmup_steps": 0,
|
| 244 |
+
"predictor_lr": 0.000425,
|
| 245 |
+
"prefetch_factor": 4,
|
| 246 |
+
"prepare_lang_cache_only": false,
|
| 247 |
+
"proprio_emb_dim": 128,
|
| 248 |
+
"recon": {
|
| 249 |
+
"enabled": false,
|
| 250 |
+
"params": {
|
| 251 |
+
"depth": 48,
|
| 252 |
+
"kernel_size": 5,
|
| 253 |
+
"mults": [
|
| 254 |
+
2,
|
| 255 |
+
3,
|
| 256 |
+
4,
|
| 257 |
+
4
|
| 258 |
+
]
|
| 259 |
+
}
|
| 260 |
+
},
|
| 261 |
+
"recon_image_weight": 1.0,
|
| 262 |
+
"recon_lang_weight": 0.1,
|
| 263 |
+
"recon_obs": false,
|
| 264 |
+
"recon_proprio_weight": 1.0,
|
| 265 |
+
"resume": null,
|
| 266 |
+
"resume_final_lr": null,
|
| 267 |
+
"resume_lr_schedule": "checkpoint",
|
| 268 |
+
"robomimic_src": "${EXTERNAL_ROOT}/asset_977d654eb1/robomimic",
|
| 269 |
+
"save_every": 28000,
|
| 270 |
+
"save_holdout_traces": true,
|
| 271 |
+
"save_latest_checkpoint": true,
|
| 272 |
+
"scaling_action_head_layers": 2,
|
| 273 |
+
"scaling_action_head_width": 192,
|
| 274 |
+
"scaling_analysis_suites": [
|
| 275 |
+
"world_tokens_per_frame_ablation"
|
| 276 |
+
],
|
| 277 |
+
"scaling_backbone_layers": 4,
|
| 278 |
+
"scaling_encoder_layers": 2,
|
| 279 |
+
"scaling_expected_world_size": 4,
|
| 280 |
+
"scaling_family_width": 768,
|
| 281 |
+
"scaling_global_batch_size": 192,
|
| 282 |
+
"scaling_grid_data_demos_per_task": 2900,
|
| 283 |
+
"scaling_grid_model_size": "n2",
|
| 284 |
+
"scaling_grid_protocol": "multitoken_d2900_n2_v1",
|
| 285 |
+
"scaling_lr_action_head": 0.0003,
|
| 286 |
+
"scaling_lr_encoder": 0.000425,
|
| 287 |
+
"scaling_lr_lookup_model": "n2",
|
| 288 |
+
"scaling_lr_lookup_status": "confirmed",
|
| 289 |
+
"scaling_lr_param_groups": [
|
| 290 |
+
"encoder",
|
| 291 |
+
"predictor",
|
| 292 |
+
"action_head"
|
| 293 |
+
],
|
| 294 |
+
"scaling_lr_predictor": 0.000425,
|
| 295 |
+
"scaling_lr_rule": "discrete_n_lookup_v20260711",
|
| 296 |
+
"scaling_lr_table_version": "n_lookup_v20260712_n4_srcenter",
|
| 297 |
+
"scaling_model_axis": "retained_world_tokens_per_frame",
|
| 298 |
+
"scaling_model_family": "n2_multitoken_interface_v1",
|
| 299 |
+
"scaling_model_size": "n2",
|
| 300 |
+
"scaling_per_rank_microbatch": 48,
|
| 301 |
+
"scaling_primary_holdout_metric": "task_macro/action_ddpm_loss",
|
| 302 |
+
"scaling_recipe": "multitoken_causal_nodyn",
|
| 303 |
+
"scaling_run_name": "multitoken_d2900_n2_k4_nodyn_seed0_280k",
|
| 304 |
+
"scaling_secondary_holdout_metric": "task_macro/action_rmse/stochastic/full10",
|
| 305 |
+
"scaling_suite": "world_tokens_per_frame_ablation",
|
| 306 |
+
"seed": 0,
|
| 307 |
+
"seq_len": 10,
|
| 308 |
+
"sequence_model": {
|
| 309 |
+
"backbone_type": "qwen2",
|
| 310 |
+
"hidden_dim": 768,
|
| 311 |
+
"params": {
|
| 312 |
+
"attn_impl": "sdpa",
|
| 313 |
+
"dropout": 0.1,
|
| 314 |
+
"ffn_hidden_size": 2048,
|
| 315 |
+
"frame_readout_index": -1,
|
| 316 |
+
"input_norm": false,
|
| 317 |
+
"max_context_len": 1024,
|
| 318 |
+
"model_dtype": "float32",
|
| 319 |
+
"n_heads": 6,
|
| 320 |
+
"n_kv_heads": null,
|
| 321 |
+
"n_layers": 4,
|
| 322 |
+
"tokens_per_frame": 4
|
| 323 |
+
},
|
| 324 |
+
"type": "frame_major_continuous_transformer"
|
| 325 |
+
},
|
| 326 |
+
"use_proprio": true,
|
| 327 |
+
"warmup_steps": 14000,
|
| 328 |
+
"weight_decay": 0.0,
|
| 329 |
+
"z_bottleneck": {
|
| 330 |
+
"dim": null
|
| 331 |
+
},
|
| 332 |
+
"eval_window_spec": "${CODE_ROOT}/experiments/common/eval_windows/c10_01.json.gz"
|
| 333 |
+
}
|
05_token_interface/multitoken_n2_k4_d2900_seed0/run_info.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"run_name": "multitoken_n2_k4_d2900_seed0",
|
| 3 |
+
"original_run_name": "multitoken_d2900_n2_k4_nodyn_seed0_280k_cuda4-5-6-7",
|
| 4 |
+
"paper_section": 5,
|
| 5 |
+
"parameters": {
|
| 6 |
+
"D": 2900,
|
| 7 |
+
"N": 2,
|
| 8 |
+
"K": 4,
|
| 9 |
+
"C_train": 10,
|
| 10 |
+
"seed": 0
|
| 11 |
+
},
|
| 12 |
+
"final_training_step": 280000
|
| 13 |
+
}
|