nvscottreed commited on
Commit
2931aaa
·
verified ·
1 Parent(s): a36bfcb

Upload folder using huggingface_hub

Browse files
SUCCESS ADDED
File without changes
config.json ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "action_horizon": 40,
3
+ "add_pos_embed": true,
4
+ "apply_sincos_state_encoding": false,
5
+ "architectures": [
6
+ "Gr00tN1d7"
7
+ ],
8
+ "attn_dropout": 0.2,
9
+ "attn_implementation": null,
10
+ "backbone_embedding_dim": 2048,
11
+ "color_jitter_params": {
12
+ "brightness": 0.3,
13
+ "contrast": 0.4,
14
+ "hue": 0.08,
15
+ "saturation": 0.5
16
+ },
17
+ "crop_fraction": 0.95,
18
+ "diffusion_model_cfg": {
19
+ "attention_head_dim": 48,
20
+ "dropout": 0.2,
21
+ "final_dropout": true,
22
+ "interleave_self_attention": true,
23
+ "norm_type": "ada_norm",
24
+ "num_attention_heads": 32,
25
+ "num_layers": 32,
26
+ "output_dim": 1024,
27
+ "positional_embeddings": null
28
+ },
29
+ "dtype": "bfloat16",
30
+ "exclude_state": false,
31
+ "formalize_language": true,
32
+ "hidden_size": 1024,
33
+ "image_crop_size": [
34
+ 230,
35
+ 230
36
+ ],
37
+ "image_target_size": [
38
+ 256,
39
+ 256
40
+ ],
41
+ "letter_box_transform": false,
42
+ "load_bf16": true,
43
+ "max_action_dim": 132,
44
+ "max_num_embodiments": 32,
45
+ "max_seq_len": 1024,
46
+ "max_state_dim": 132,
47
+ "model_dtype": "bfloat16",
48
+ "model_type": "Gr00tN1d7",
49
+ "noise_beta_alpha": 1.5,
50
+ "noise_beta_beta": 1.0,
51
+ "noise_s": 0.999,
52
+ "num_inference_timesteps": 4,
53
+ "num_timestep_buckets": 1000,
54
+ "random_history_crop": true,
55
+ "random_rotation_angle": 0,
56
+ "reproject_vision": false,
57
+ "rtc_ramp_rate": 6.0,
58
+ "select_layer": 16,
59
+ "shortest_image_edge": 256,
60
+ "state_dropout_prob": 0.2,
61
+ "state_gaussian_noise_std": 0.0,
62
+ "transformers_version": "4.57.1",
63
+ "tune_diffusion_model": true,
64
+ "tune_linear": true,
65
+ "tune_llm": true,
66
+ "tune_projector": true,
67
+ "tune_top_llm_layers": 0,
68
+ "tune_visual": true,
69
+ "tune_vlln": true,
70
+ "use_albumentations": true,
71
+ "use_alternate_vl_dit": true,
72
+ "use_flash_attention": true,
73
+ "use_future_tokens": false,
74
+ "use_mean_std": false,
75
+ "use_percentiles": true,
76
+ "use_vl_self_attention": true,
77
+ "use_vlln": true,
78
+ "vl_self_attention_cfg": {
79
+ "attention_head_dim": 64,
80
+ "dropout": 0.2,
81
+ "final_dropout": true,
82
+ "num_attention_heads": 32,
83
+ "num_layers": 4,
84
+ "positional_embeddings": null
85
+ },
86
+ "model_name": "nvidia/Cosmos-Reason2-2B"
87
+ }
embodiment_id.json ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "robocasa_panda_omron": 13,
3
+ "oxe_droid": 17,
4
+ "oxe_fractal": 18,
5
+ "oxe_language_table": 19,
6
+ "oxe_bridge": 20,
7
+ "unknown": 22,
8
+ "gr1_unified": 20,
9
+ "agibot": 26,
10
+ "sim_behavior_r1_pro": 23,
11
+ "xdof": 24,
12
+ "xdof_oss_data": 25,
13
+ "unitree_g1_full_body_with_waist_height_nav_cmd": 25,
14
+ "real_r1_pro_sharpa": 27,
15
+ "real_r1_pro_sharpa_add_view": 27,
16
+ "real_r1_pro_sharpa_relative_arm_joint": 26,
17
+ "real_r1_pro_sharpa_delta_eef": 26,
18
+ "real_r1_pro_sharpa_absolute_eef": 26,
19
+ "real_r1_pro_sharpa_meanstd": 26,
20
+ "real_r1_pro_sharpa_relative_eef": 26,
21
+ "real_r1_pro_sharpa_relative_eef_add_view": 26,
22
+ "real_r1_pro_sharpa_relative_eef_relative_hand": 26,
23
+ "real_r1_pro_sharpa_relative_eef_human": 26,
24
+ "real_r1_pro_sharpa_relative_eef_human_add_view": 26,
25
+ "real_r1_pro_sharpa_relative_eef_human_relative_hand": 26,
26
+ "real_r1_pro_sharpa_relative_eef_egodex": 26,
27
+ "real_r1_pro_sharpa_relative_eef_egodex_relative_hand": 26,
28
+ "real_r1_pro_sharpa_relative_eef_egodex_wrist_only": 26,
29
+ "real_r1_pro_sharpa_relative_eef_maxinsights": 26,
30
+ "real_r1_pro_sharpa_relative_eef_maxinsights_relative_hand": 26,
31
+ "real_r1_pro_sharpa_relative_eef_mecka": 26,
32
+ "real_r1_pro_sharpa_relative_eef_mecka_relative_hand": 26,
33
+ "real_g1_relative_eef_absolute_joints": 25,
34
+ "real_g1_relative_eef_absolute_joints_wrist_cam": 25,
35
+ "real_g1_relative_eef_relative_joints": 25,
36
+ "real_r1_pro_sharpa_relative_eef_relative_hand_relative_joint": 26,
37
+ "real_r1_pro_sharpa_relative_joint": 29,
38
+ "oxe_droid_relative_eef_relative_joint": 24,
39
+ "oxe_droid_relative_eef_relative_joint_swapped": 24,
40
+ "oxe_droid_relative_eef_relative_joint_upweight_z": 24,
41
+ "oxe_droid_relative_eef_relative_joint_upweight_z_swapped": 24,
42
+ "oxe_droid_relative_eef_relative_joint_3view": 24,
43
+ "oxe_droid_relative_eef_relative_joint_3view_swapped": 24,
44
+ "oxe_droid_relative_eef": 24,
45
+ "oxe_droid_joint_position_relative": 24,
46
+ "xdof_relative_eef_relative_joint": 27,
47
+ "xdof_relative_eef_relative_joint_subtask": 27,
48
+ "xdof_relative_eef": 27,
49
+ "xdof_relative_joint": 28,
50
+ "simpler_env_google": 0,
51
+ "simpler_env_widowx": 1,
52
+ "libero_sim": 2,
53
+ "droid_sim": 3,
54
+ "xdof_relative_eef_subtask": 27,
55
+ "xdof_relative_joint_subtask": 28
56
+ }
experiment_cfg/conf.yaml ADDED
@@ -0,0 +1,373 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ load_config_path: groot/vla/omni/configs/experiments/droid_relativeeef_relativejoint_upweight_z.yaml
2
+ model:
3
+ return_dict: true
4
+ output_hidden_states: false
5
+ torchscript: false
6
+ dtype: null
7
+ pruned_heads: {}
8
+ tie_word_embeddings: true
9
+ chunk_size_feed_forward: 0
10
+ is_encoder_decoder: false
11
+ is_decoder: false
12
+ cross_attention_hidden_size: null
13
+ add_cross_attention: false
14
+ tie_encoder_decoder: false
15
+ architectures: null
16
+ finetuning_task: null
17
+ id2label:
18
+ 0: LABEL_0
19
+ 1: LABEL_1
20
+ label2id:
21
+ LABEL_0: 0
22
+ LABEL_1: 1
23
+ task_specific_params: null
24
+ problem_type: null
25
+ tokenizer_class: null
26
+ prefix: null
27
+ bos_token_id: null
28
+ pad_token_id: null
29
+ eos_token_id: null
30
+ sep_token_id: null
31
+ decoder_start_token_id: null
32
+ max_length: 20
33
+ min_length: 0
34
+ do_sample: false
35
+ early_stopping: false
36
+ num_beams: 1
37
+ temperature: 1.0
38
+ top_k: 50
39
+ top_p: 1.0
40
+ typical_p: 1.0
41
+ repetition_penalty: 1.0
42
+ length_penalty: 1.0
43
+ no_repeat_ngram_size: 0
44
+ encoder_no_repeat_ngram_size: 0
45
+ bad_words_ids: null
46
+ num_return_sequences: 1
47
+ output_scores: false
48
+ return_dict_in_generate: false
49
+ forced_bos_token_id: null
50
+ forced_eos_token_id: null
51
+ remove_invalid_values: false
52
+ exponential_decay_length_penalty: null
53
+ suppress_tokens: null
54
+ begin_suppress_tokens: null
55
+ num_beam_groups: 1
56
+ diversity_penalty: 0.0
57
+ transformers_version: null
58
+ model_type: GrootN1d5Qwen
59
+ model_dtype: bfloat16
60
+ vlm_backend: qwen3
61
+ vlm_model_path: nvidia/Cosmos-Reason2-2B
62
+ backbone_embedding_dim: 2048
63
+ tune_llm: true
64
+ tune_top_llm_layers: 0
65
+ tune_visual: true
66
+ tune_linear: true
67
+ select_layer: 16
68
+ reproject_vision: false
69
+ use_flash_attention: true
70
+ load_bf16: true
71
+ exclude_state: false
72
+ image_crop_size:
73
+ - 230
74
+ - 230
75
+ image_target_size:
76
+ - 256
77
+ - 256
78
+ random_rotation_angle: 0
79
+ color_jitter_params:
80
+ brightness: 0.3
81
+ contrast: 0.4
82
+ saturation: 0.5
83
+ hue: 0.08
84
+ formalize_language: true
85
+ action_space_prompt: false
86
+ apply_sincos_state_encoding: false
87
+ letter_box_transform: false
88
+ use_percentiles: true
89
+ use_mean_std: false
90
+ use_albumentations: true
91
+ shortest_image_edge: 256
92
+ crop_fraction: 0.95
93
+ random_history_crop: true
94
+ state_gaussian_noise_std: 0.0
95
+ do_human_interpolation: false
96
+ interpolation_steps: 20
97
+ human_embodiment_tags: null
98
+ max_state_dim: 132
99
+ max_action_dim: 132
100
+ action_horizon: 40
101
+ hidden_size: 1024
102
+ dit_latent_dim: 1536
103
+ state_dropout_prob: 0.5
104
+ language_dropout_prob: 0.0
105
+ add_pos_embed: true
106
+ attn_dropout: 0.2
107
+ use_vlln: true
108
+ use_vl_self_attention: true
109
+ max_seq_len: 1024
110
+ use_future_tokens: false
111
+ use_alternate_vl_dit: true
112
+ vl_self_attention_cfg:
113
+ positional_embeddings: null
114
+ num_layers: 4
115
+ num_attention_heads: 32
116
+ attention_head_dim: 64
117
+ dropout: 0.2
118
+ final_dropout: true
119
+ diffusion_model_cfg:
120
+ positional_embeddings: null
121
+ num_layers: 32
122
+ num_attention_heads: 32
123
+ attention_head_dim: 48
124
+ norm_type: ada_norm
125
+ dropout: 0.2
126
+ final_dropout: true
127
+ output_dim: 1024
128
+ interleave_self_attention: true
129
+ cross_attention_dim: 2048
130
+ num_inference_timesteps: 4
131
+ noise_beta_alpha: 1.5
132
+ noise_beta_beta: 1.0
133
+ noise_s: 0.999
134
+ num_timestep_buckets: 1000
135
+ tune_projector: true
136
+ tune_diffusion_model: true
137
+ tune_vlln: true
138
+ max_num_embodiments: 32
139
+ unified_embodiment_id: false
140
+ rtc_ramp_rate: 6.0
141
+ tf_legacy_loss: false
142
+ use_bfloat16: false
143
+ data:
144
+ datasets:
145
+ - dataset_paths:
146
+ - /mnt/aws-lfs-02/shared/datasets/roid_101_success_idlefiltered_n17
147
+ embodiment_tag: oxe_droid_relative_eef_relative_joint
148
+ mix_ratio: 1.0
149
+ dataset_type: physical_embodiment
150
+ - dataset_paths:
151
+ - /mnt/aws-lfs-02/shared/datasets/droid_101_success_idlefiltered_n17
152
+ embodiment_tag: oxe_droid_relative_eef_relative_joint_swapped
153
+ mix_ratio: 1.0
154
+ dataset_type: physical_embodiment
155
+ modality_configs:
156
+ oxe_droid_relative_eef_relative_joint:
157
+ video:
158
+ delta_indices:
159
+ - 0
160
+ modality_keys:
161
+ - exterior_image_1_left
162
+ - wrist_image_left
163
+ normalization_mode: null
164
+ action_representation: null
165
+ exclude_state: false
166
+ action_type: null
167
+ action_format: null
168
+ normalize_rotation: true
169
+ wrist_keys: null
170
+ hand_keys: null
171
+ extra_keys: null
172
+ loss_weights: null
173
+ state:
174
+ delta_indices:
175
+ - 0
176
+ modality_keys:
177
+ - eef_9d
178
+ - gripper_position
179
+ - joint_position
180
+ normalization_mode: null
181
+ action_representation: null
182
+ exclude_state: false
183
+ action_type: null
184
+ action_format: null
185
+ normalize_rotation: true
186
+ wrist_keys: null
187
+ hand_keys: null
188
+ extra_keys: null
189
+ loss_weights: null
190
+ action:
191
+ delta_indices:
192
+ - 0
193
+ - 1
194
+ - 2
195
+ - 3
196
+ - 4
197
+ - 5
198
+ - 6
199
+ - 7
200
+ - 8
201
+ - 9
202
+ - 10
203
+ - 11
204
+ - 12
205
+ - 13
206
+ - 14
207
+ - 15
208
+ - 16
209
+ - 17
210
+ - 18
211
+ - 19
212
+ - 20
213
+ - 21
214
+ - 22
215
+ - 23
216
+ - 24
217
+ - 25
218
+ - 26
219
+ - 27
220
+ - 28
221
+ - 29
222
+ - 30
223
+ - 31
224
+ - 32
225
+ - 33
226
+ - 34
227
+ - 35
228
+ - 36
229
+ - 37
230
+ - 38
231
+ - 39
232
+ modality_keys:
233
+ - eef_9d
234
+ - gripper_position
235
+ - joint_position
236
+ normalization_mode: null
237
+ action_representation:
238
+ - {}
239
+ - {}
240
+ - {}
241
+ exclude_state: false
242
+ action_type:
243
+ - {}
244
+ - {}
245
+ - {}
246
+ action_format:
247
+ - {}
248
+ - {}
249
+ - {}
250
+ normalize_rotation: true
251
+ wrist_keys:
252
+ - eef_9d
253
+ hand_keys:
254
+ - gripper_position
255
+ extra_keys:
256
+ - joint_position
257
+ loss_weights:
258
+ eef_9d:
259
+ - 1.0
260
+ - 1.0
261
+ - 2.0
262
+ - 1.0
263
+ - 1.0
264
+ - 1.0
265
+ - 1.0
266
+ - 1.0
267
+ - 1.0
268
+ gripper_position:
269
+ - 2.0
270
+ joint_position:
271
+ - 1.0
272
+ - 2.0
273
+ - 1.0
274
+ - 2.0
275
+ - 1.0
276
+ - 1.0
277
+ - 1.0
278
+ language:
279
+ delta_indices:
280
+ - 0
281
+ modality_keys:
282
+ - annotation.language.language_instruction
283
+ - annotation.language.language_instruction_2
284
+ - annotation.language.language_instruction_3
285
+ - annotation.language.language_instruction_4
286
+ - annotation.language.language_instruction_5
287
+ - annotation.language.language_instruction_6
288
+ normalization_mode: null
289
+ action_representation: null
290
+ exclude_state: false
291
+ action_type: null
292
+ action_format: null
293
+ normalize_rotation: true
294
+ wrist_keys: null
295
+ hand_keys: null
296
+ extra_keys: null
297
+ loss_weights: null
298
+ download_cache: false
299
+ shard_size: 1024
300
+ episode_sampling_rate: 0.1
301
+ num_shards_per_epoch: 100000
302
+ override_pretraining_statistics: false
303
+ mode: single_turn
304
+ random_chop: 0.0
305
+ mock_dataset_mode: false
306
+ num_prompt_trajectories: 2
307
+ variable_num_demos: false
308
+ max_prompt_trajectories: 5
309
+ shuffle: true
310
+ seed: 24
311
+ subsample_ratio: 1.0
312
+ image_crop_size:
313
+ - 244
314
+ - 244
315
+ image_target_size:
316
+ - 224
317
+ - 224
318
+ video_backend: torchcodec
319
+ training:
320
+ output_dir: /mnt/aws-lfs-02/shared/ckpts/ruijiez/galaxea_sharpa/n17_ph2_droid_posttrain/posttrain_droid_qwen3vl_2b_finetuned_double_dit_ckpt130k_upweight_z_unfrozen_vlm_wsd_lr3e-5_100ksteps_state_dropout50
321
+ experiment_name: null
322
+ max_steps: 100000
323
+ global_batch_size: 1024
324
+ batch_size: 32
325
+ gradient_accumulation_steps: 1
326
+ use_muon: false
327
+ muon_lr: 0.005
328
+ use_legacy_wd_application: false
329
+ learning_rate: 3.0e-05
330
+ lr_scheduler_type: wsd
331
+ weight_decay: 1.0e-05
332
+ warmup_ratio: 0.0
333
+ warmup_steps: 0
334
+ max_grad_norm: 1.0
335
+ wsd_stable_ratio: 0.8
336
+ wsd_decay_type: cosine
337
+ optim: adamw_torch_fused
338
+ start_from_checkpoint: /mnt/aws-lfs-02/shared/ckpts/ruijiez/galaxea_sharpa/stable_checkpoints/qwen3vl_midtrain_ckpt130k
339
+ tf32: true
340
+ fp16: false
341
+ bf16: true
342
+ eval_bf16: true
343
+ logging_steps: 10
344
+ save_steps: 1000
345
+ save_total_limit: 5
346
+ save_vl_model: false
347
+ upload_checkpoints: true
348
+ upload_every: 1000
349
+ upload_last_n_checkpoints: 5
350
+ max_concurrent_uploads: 2
351
+ eval_strategy: 'no'
352
+ eval_steps: 500
353
+ eval_set_split_ratio: 0.1
354
+ eval_batch_size: 2
355
+ save_best_eval_metric_name: ''
356
+ save_best_eval_metric_greater_is_better: true
357
+ deepspeed_stage: 2
358
+ gradient_checkpointing: false
359
+ use_ddp: false
360
+ num_gpus: 64
361
+ dataloader_num_workers: 4
362
+ remove_unused_columns: false
363
+ use_wandb: true
364
+ wandb_project: n17_ph2_droid_posttrain
365
+ enable_profiling: false
366
+ max_retries: 3
367
+ skip_spike: true
368
+ skip_spike_threshold: 5.0
369
+ skip_spike_ema_alpha: 0.99
370
+ skip_spike_max_consecutive: 10
371
+ assert_loss_less_than: null
372
+ max_steps: 100000
373
+ save_steps: 1000
experiment_cfg/config.yaml ADDED
@@ -0,0 +1,382 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ !!python/object:groot.vla.omni.configs.base_config.Config
2
+ data: !!python/object:groot.vla.omni.configs.data.data_config.DataConfig
3
+ datasets:
4
+ - !!python/object:groot.vla.omni.configs.data.data_config.SingleDatasetConfig
5
+ dataset_paths:
6
+ - /mnt/aws-lfs-02/shared/datasets/droid_101_success_idlefiltered_n17
7
+ dataset_type: physical_embodiment
8
+ embodiment_tag: oxe_droid_relative_eef_relative_joint
9
+ mix_ratio: 1.0
10
+ - !!python/object:groot.vla.omni.configs.data.data_config.SingleDatasetConfig
11
+ dataset_paths:
12
+ - /mnt/aws-lfs-02/shared/datasets/droid_101_success_idlefiltered_n17
13
+ dataset_type: physical_embodiment
14
+ embodiment_tag: oxe_droid_relative_eef_relative_joint
15
+ mix_ratio: 1.0
16
+ download_cache: false
17
+ episode_sampling_rate: 0.1
18
+ image_crop_size:
19
+ - 244
20
+ - 244
21
+ image_target_size:
22
+ - 224
23
+ - 224
24
+ max_prompt_trajectories: 5
25
+ mock_dataset_mode: false
26
+ modality_configs:
27
+ oxe_droid_relative_eef_relative_joint:
28
+ action: !!python/object:groot.vla.omni.data.types.ModalityConfig
29
+ action_format:
30
+ - &id004 !!python/object/apply:groot.vla.omni.data.types.ActionFormat
31
+ - xyz+rot6d
32
+ - &id001 !!python/object/apply:groot.vla.omni.data.types.ActionFormat
33
+ - default
34
+ - *id001
35
+ action_representation:
36
+ - &id002 !!python/object/apply:groot.vla.omni.data.types.ActionRepresentation
37
+ - relative
38
+ - &id005 !!python/object/apply:groot.vla.omni.data.types.ActionRepresentation
39
+ - absolute
40
+ - *id002
41
+ action_type:
42
+ - &id006 !!python/object/apply:groot.vla.omni.data.types.ActionType
43
+ - eef
44
+ - &id003 !!python/object/apply:groot.vla.omni.data.types.ActionType
45
+ - non_eef
46
+ - *id003
47
+ delta_indices:
48
+ - 0
49
+ - 1
50
+ - 2
51
+ - 3
52
+ - 4
53
+ - 5
54
+ - 6
55
+ - 7
56
+ - 8
57
+ - 9
58
+ - 10
59
+ - 11
60
+ - 12
61
+ - 13
62
+ - 14
63
+ - 15
64
+ - 16
65
+ - 17
66
+ - 18
67
+ - 19
68
+ - 20
69
+ - 21
70
+ - 22
71
+ - 23
72
+ - 24
73
+ - 25
74
+ - 26
75
+ - 27
76
+ - 28
77
+ - 29
78
+ - 30
79
+ - 31
80
+ - 32
81
+ - 33
82
+ - 34
83
+ - 35
84
+ - 36
85
+ - 37
86
+ - 38
87
+ - 39
88
+ exclude_state: false
89
+ extra_keys:
90
+ - joint_position
91
+ hand_keys:
92
+ - gripper_position
93
+ loss_weights:
94
+ eef_9d:
95
+ - 1.0
96
+ - 1.0
97
+ - 2.0
98
+ - 1.0
99
+ - 1.0
100
+ - 1.0
101
+ - 1.0
102
+ - 1.0
103
+ - 1.0
104
+ gripper_position:
105
+ - 2.0
106
+ joint_position:
107
+ - 1.0
108
+ - 2.0
109
+ - 1.0
110
+ - 2.0
111
+ - 1.0
112
+ - 1.0
113
+ - 1.0
114
+ modality_keys:
115
+ - eef_9d
116
+ - gripper_position
117
+ - joint_position
118
+ normalization_mode: null
119
+ normalize_rotation: true
120
+ wrist_keys:
121
+ - eef_9d
122
+ language: !!python/object:groot.vla.omni.data.types.ModalityConfig
123
+ action_format: null
124
+ action_representation: null
125
+ action_type: null
126
+ delta_indices:
127
+ - 0
128
+ exclude_state: false
129
+ extra_keys: null
130
+ hand_keys: null
131
+ loss_weights: null
132
+ modality_keys:
133
+ - annotation.language.language_instruction
134
+ - annotation.language.language_instruction_2
135
+ - annotation.language.language_instruction_3
136
+ normalization_mode: null
137
+ normalize_rotation: true
138
+ wrist_keys: null
139
+ state: !!python/object:groot.vla.omni.data.types.ModalityConfig
140
+ action_format: null
141
+ action_representation: null
142
+ action_type: null
143
+ delta_indices:
144
+ - 0
145
+ exclude_state: false
146
+ extra_keys: null
147
+ hand_keys: null
148
+ loss_weights: null
149
+ modality_keys:
150
+ - eef_9d
151
+ - gripper_position
152
+ - joint_position
153
+ normalization_mode: null
154
+ normalize_rotation: true
155
+ wrist_keys: null
156
+ video: !!python/object:groot.vla.omni.data.types.ModalityConfig
157
+ action_format: null
158
+ action_representation: null
159
+ action_type: null
160
+ delta_indices:
161
+ - -15
162
+ - 0
163
+ exclude_state: false
164
+ extra_keys: null
165
+ hand_keys: null
166
+ loss_weights: null
167
+ modality_keys:
168
+ - exterior_image_1_left
169
+ - wrist_image_left
170
+ normalization_mode: null
171
+ normalize_rotation: true
172
+ wrist_keys: null
173
+ mode: single_turn
174
+ num_prompt_trajectories: 2
175
+ num_shards_per_epoch: 100000
176
+ override_pretraining_statistics: false
177
+ random_chop: 0.0
178
+ seed: 24
179
+ shard_size: 1024
180
+ shuffle: true
181
+ subsample_ratio: 1.0
182
+ variable_num_demos: false
183
+ video_backend: torchcodec
184
+ load_config_path: groot/vla/omni/configs/experiments/droid_relativeeef_relativejoint_upweight_z.yaml
185
+ model: !!python/object:groot.vla.omni.configs.model.groot_n1d5_qwen.GrootN1d5QwenConfig
186
+ _attn_implementation_internal: null
187
+ _commit_hash: null
188
+ _name_or_path: ''
189
+ _output_attentions: false
190
+ action_horizon: 40
191
+ action_space_prompt: false
192
+ add_cross_attention: false
193
+ add_pos_embed: true
194
+ apply_sincos_state_encoding: false
195
+ architectures: null
196
+ attn_dropout: 0.2
197
+ backbone_embedding_dim: 2048
198
+ bad_words_ids: null
199
+ begin_suppress_tokens: null
200
+ bos_token_id: null
201
+ chunk_size_feed_forward: 0
202
+ color_jitter_params:
203
+ brightness: 0.3
204
+ contrast: 0.4
205
+ hue: 0.08
206
+ saturation: 0.5
207
+ crop_fraction: 0.95
208
+ cross_attention_hidden_size: null
209
+ decoder_start_token_id: null
210
+ diffusion_model_cfg:
211
+ attention_head_dim: 48
212
+ cross_attention_dim: 2048
213
+ dropout: 0.2
214
+ final_dropout: true
215
+ interleave_self_attention: true
216
+ norm_type: ada_norm
217
+ num_attention_heads: 32
218
+ num_layers: 32
219
+ output_dim: 1024
220
+ positional_embeddings: null
221
+ dit_latent_dim: 1536
222
+ diversity_penalty: 0.0
223
+ do_human_interpolation: false
224
+ do_sample: false
225
+ dtype: null
226
+ early_stopping: false
227
+ encoder_no_repeat_ngram_size: 0
228
+ eos_token_id: null
229
+ exclude_state: false
230
+ exponential_decay_length_penalty: null
231
+ finetuning_task: null
232
+ forced_bos_token_id: null
233
+ forced_eos_token_id: null
234
+ formalize_language: true
235
+ hidden_size: 1024
236
+ human_embodiment_tags: null
237
+ id2label:
238
+ 0: LABEL_0
239
+ 1: LABEL_1
240
+ image_crop_size: !!python/tuple
241
+ - 230
242
+ - 230
243
+ image_target_size: !!python/tuple
244
+ - 256
245
+ - 256
246
+ interpolation_steps: 20
247
+ is_decoder: false
248
+ is_encoder_decoder: false
249
+ label2id:
250
+ LABEL_0: 0
251
+ LABEL_1: 1
252
+ language_dropout_prob: 0.0
253
+ length_penalty: 1.0
254
+ letter_box_transform: false
255
+ load_bf16: true
256
+ max_action_dim: 132
257
+ max_length: 20
258
+ max_num_embodiments: 32
259
+ max_seq_len: 1024
260
+ max_state_dim: 132
261
+ min_length: 0
262
+ model_dtype: bfloat16
263
+ model_type: GrootN1d5Qwen
264
+ no_repeat_ngram_size: 0
265
+ noise_beta_alpha: 1.5
266
+ noise_beta_beta: 1.0
267
+ noise_s: 0.999
268
+ num_beam_groups: 1
269
+ num_beams: 1
270
+ num_inference_timesteps: 4
271
+ num_return_sequences: 1
272
+ num_timestep_buckets: 1000
273
+ output_hidden_states: false
274
+ output_scores: false
275
+ pad_token_id: null
276
+ prefix: null
277
+ problem_type: null
278
+ pruned_heads: {}
279
+ random_history_crop: true
280
+ random_rotation_angle: 0
281
+ remove_invalid_values: false
282
+ repetition_penalty: 1.0
283
+ reproject_vision: false
284
+ return_dict: true
285
+ return_dict_in_generate: false
286
+ rtc_ramp_rate: 6.0
287
+ select_layer: 16
288
+ sep_token_id: null
289
+ shortest_image_edge: 256
290
+ state_dropout_prob: 0.5
291
+ state_gaussian_noise_std: 0.0
292
+ suppress_tokens: null
293
+ task_specific_params: null
294
+ temperature: 1.0
295
+ tf_legacy_loss: false
296
+ tie_encoder_decoder: false
297
+ tie_word_embeddings: true
298
+ tokenizer_class: null
299
+ top_k: 50
300
+ top_p: 1.0
301
+ torchscript: false
302
+ transformers_version: null
303
+ tune_diffusion_model: true
304
+ tune_linear: true
305
+ tune_llm: true
306
+ tune_projector: true
307
+ tune_top_llm_layers: 0
308
+ tune_visual: true
309
+ tune_vlln: true
310
+ typical_p: 1.0
311
+ unified_embodiment_id: false
312
+ use_albumentations: true
313
+ use_alternate_vl_dit: true
314
+ use_bfloat16: false
315
+ use_flash_attention: true
316
+ use_future_tokens: false
317
+ use_mean_std: false
318
+ use_percentiles: true
319
+ use_vl_self_attention: true
320
+ use_vlln: true
321
+ vl_self_attention_cfg:
322
+ attention_head_dim: 64
323
+ dropout: 0.2
324
+ final_dropout: true
325
+ num_attention_heads: 32
326
+ num_layers: 4
327
+ positional_embeddings: null
328
+ vlm_backend: qwen3
329
+ vlm_model_path: nvidia/Cosmos-Reason2-2B
330
+ training: !!python/object:groot.vla.omni.configs.training.training_config.TrainingConfig
331
+ assert_loss_less_than: null
332
+ batch_size: 32
333
+ bf16: true
334
+ dataloader_num_workers: 4
335
+ deepspeed_stage: 2
336
+ enable_profiling: false
337
+ eval_batch_size: 2
338
+ eval_bf16: true
339
+ eval_set_split_ratio: 0.1
340
+ eval_steps: 500
341
+ eval_strategy: 'no'
342
+ experiment_name: null
343
+ fp16: false
344
+ global_batch_size: 1024
345
+ gradient_accumulation_steps: 1
346
+ gradient_checkpointing: false
347
+ learning_rate: 3.0e-05
348
+ logging_steps: 10
349
+ lr_scheduler_type: wsd
350
+ max_concurrent_uploads: 2
351
+ max_grad_norm: 1.0
352
+ max_retries: 3
353
+ max_steps: 100000
354
+ muon_lr: 0.005
355
+ num_gpus: 64
356
+ optim: adamw_torch_fused
357
+ output_dir: /mnt/aws-lfs-02/shared/ckpts/ruijiez/galaxea_sharpa/n17_ph2_droid_posttrain/posttrain_droid_qwen3vl_2b_finetuned_double_dit_ckpt130k_upweight_z_unfrozen_vlm_wsd_lr3e-5_100ksteps_state_dropout50
358
+ remove_unused_columns: false
359
+ save_best_eval_metric_greater_is_better: true
360
+ save_best_eval_metric_name: ''
361
+ save_steps: 1000
362
+ save_total_limit: 5
363
+ save_vl_model: false
364
+ skip_spike: true
365
+ skip_spike_ema_alpha: 0.99
366
+ skip_spike_max_consecutive: 10
367
+ skip_spike_threshold: 5.0
368
+ start_from_checkpoint: /mnt/aws-lfs-02/shared/ckpts/ruijiez/galaxea_sharpa/stable_checkpoints/qwen3vl_midtrain_ckpt130k
369
+ tf32: true
370
+ upload_checkpoints: true
371
+ upload_every: 1000
372
+ upload_last_n_checkpoints: 5
373
+ use_ddp: false
374
+ use_legacy_wd_application: false
375
+ use_muon: false
376
+ use_wandb: true
377
+ wandb_project: n17_ph2_droid_posttrain
378
+ warmup_ratio: 0.0
379
+ warmup_steps: 0
380
+ weight_decay: 1.0e-05
381
+ wsd_decay_type: cosine
382
+ wsd_stable_ratio: 0.8
experiment_cfg/dataset_statistics.json ADDED
The diff for this file is too large to render. See raw diff
 
latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step100000
model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:68d885c9684bb7d4781389873e4b7d33202b5618e70a83f2e78187a5fb839202
3
+ size 4990519232
model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aa4c6e553ea8454500354352368bcbb7e4f0fb32a9816b20d5b25c231f13a8fd
3
+ size 1919980184
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
processor_config.json ADDED
@@ -0,0 +1,135 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "processor_class": "Gr00tN1d7Processor",
3
+ "processor_kwargs": {
4
+ "modality_configs": {
5
+ "oxe_droid_relative_eef_relative_joint": {
6
+ "video": {
7
+ "delta_indices": [
8
+ 0
9
+ ],
10
+ "modality_keys": [
11
+ "exterior_image_1_left",
12
+ "wrist_image_left"
13
+ ]
14
+ },
15
+ "state": {
16
+ "delta_indices": [
17
+ 0
18
+ ],
19
+ "modality_keys": [
20
+ "eef_9d",
21
+ "gripper_position",
22
+ "joint_position"
23
+ ]
24
+ },
25
+ "action": {
26
+ "delta_indices": [
27
+ 0,
28
+ 1,
29
+ 2,
30
+ 3,
31
+ 4,
32
+ 5,
33
+ 6,
34
+ 7,
35
+ 8,
36
+ 9,
37
+ 10,
38
+ 11,
39
+ 12,
40
+ 13,
41
+ 14,
42
+ 15,
43
+ 16,
44
+ 17,
45
+ 18,
46
+ 19,
47
+ 20,
48
+ 21,
49
+ 22,
50
+ 23,
51
+ 24,
52
+ 25,
53
+ 26,
54
+ 27,
55
+ 28,
56
+ 29,
57
+ 30,
58
+ 31,
59
+ 32,
60
+ 33,
61
+ 34,
62
+ 35,
63
+ 36,
64
+ 37,
65
+ 38,
66
+ 39
67
+ ],
68
+ "modality_keys": [
69
+ "eef_9d",
70
+ "gripper_position",
71
+ "joint_position"
72
+ ],
73
+ "action_configs": [
74
+ {
75
+ "rep": "RELATIVE",
76
+ "type": "EEF",
77
+ "format": "XYZ_ROT6D",
78
+ "state_key": "eef_9d"
79
+ },
80
+ {
81
+ "rep": "ABSOLUTE",
82
+ "type": "NON_EEF",
83
+ "format": "DEFAULT",
84
+ "state_key": "gripper_position"
85
+ },
86
+ {
87
+ "rep": "RELATIVE",
88
+ "type": "NON_EEF",
89
+ "format": "DEFAULT",
90
+ "state_key": "joint_position"
91
+ }
92
+ ]
93
+ },
94
+ "language": {
95
+ "delta_indices": [
96
+ 0
97
+ ],
98
+ "modality_keys": [
99
+ "annotation.language.language_instruction"
100
+ ]
101
+ }
102
+ }
103
+ },
104
+ "use_percentiles": true,
105
+ "use_mean_std": false,
106
+ "image_crop_size": [
107
+ 230,
108
+ 230
109
+ ],
110
+ "image_target_size": [
111
+ 256,
112
+ 256
113
+ ],
114
+ "formalize_language": true,
115
+ "max_state_dim": 132,
116
+ "max_action_dim": 132,
117
+ "apply_sincos_state_encoding": false,
118
+ "color_jitter_params": {
119
+ "brightness": 0.3,
120
+ "contrast": 0.4,
121
+ "saturation": 0.5,
122
+ "hue": 0.08
123
+ },
124
+ "random_rotation_angle": 0,
125
+ "letter_box_transform": false,
126
+ "exclude_state": false,
127
+ "state_dropout_prob": 0.2,
128
+ "state_gaussian_noise_std": 0.0,
129
+ "use_albumentations": true,
130
+ "shortest_image_edge": 256,
131
+ "crop_fraction": 0.95,
132
+ "max_action_horizon": 40,
133
+ "use_relative_action": true
134
+ }
135
+ }
scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0e68cd5c0983910f321ec0ab989e4ea8c30670cfdbd27cf252aeb2b95587256c
3
+ size 1263
statistics.json ADDED
The diff for this file is too large to render. See raw diff
 
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8d636db8bb639e87538810ab7d76176a6a334d903728dd0bd0b3d2358f668e1b
3
+ size 8259