Hugging Face's logo

elijahgalahad
/
checkpoint-action_a_bbox_a

Model card Files
xet
 Community
elijahgalahad commited on
Commit
282a411
·
verified ·
1 Parent(s): 70320a5

Upload folder using huggingface_hub

Browse files
Files changed (5) hide show
  1. cfg.yaml +201 -0
  2. train-cfg.yaml +280 -0
  3. wandb/run-20251225_195640-1lai13bo/files/cfg.yaml +201 -0
  4. wandb/run-20251225_195640-1lai13bo/files/config.yaml +344 -0
  5. wandb/run-20251225_195640-1lai13bo/files/train-cfg.yaml +280 -0
cfg.yaml ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_workers: 4
2
+ prefetch_factor: 2
3
+ split_seed: 42
4
+ epoch_iterator: eager
5
+ epochs: 121
6
+ batch_size: 32
7
+ grad_accum_steps: 1
8
+ lr:
9
+ base: 5.0e-05
10
+ vlm_bridge: 1.0e-05
11
+ action_expert: 5.0e-05
12
+ warmup_steps: 0
13
+ cosine_anneal_epoch: 0
14
+ betas:
15
+ - 0.99
16
+ - 0.9999
17
+ eps: 1.0e-08
18
+ weight_decay: 0.0001
19
+ clip_grad_norm: 1.0
20
+ exp_name: pi-training
21
+ log_interval: 32
22
+ eval_interval: 512
23
+ save_interval: 40
24
+ data:
25
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
26
+ root_path: null
27
+ action_horizon: null
28
+ state_history: null
29
+ input_transforms: []
30
+ output_transforms: []
31
+ output_inv_transforms: []
32
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
33
+ noise_cfg: null
34
+ repo_id: horipse01/lerobot_merged_restricted
35
+ bbox_only: false
36
+ remove_bbox: false
37
+ episodes: null
38
+ splits:
39
+ - .*
40
+ train_data:
41
+ datasets:
42
+ action_a:
43
+ data:
44
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
45
+ root_path: null
46
+ action_horizon: null
47
+ state_history: null
48
+ input_transforms: []
49
+ output_transforms: []
50
+ output_inv_transforms: []
51
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
52
+ noise_cfg: null
53
+ repo_id: horipse01/lerobot_merged_restricted
54
+ bbox_only: false
55
+ remove_bbox: false
56
+ episodes: null
57
+ splits:
58
+ - .*
59
+ batch_size: 64
60
+ eval_data:
61
+ datasets:
62
+ action_train:
63
+ data:
64
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
65
+ root_path: null
66
+ action_horizon: null
67
+ state_history: null
68
+ input_transforms: []
69
+ output_transforms: []
70
+ output_inv_transforms: []
71
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
72
+ noise_cfg: null
73
+ repo_id: horipse01/lerobot_merged_restricted
74
+ bbox_only: false
75
+ remove_bbox: false
76
+ episodes: null
77
+ splits:
78
+ - .*
79
+ eval_fraction: 0.02
80
+ eval_type: sample_mse
81
+ action_test:
82
+ data:
83
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
84
+ root_path: null
85
+ action_horizon: null
86
+ state_history: null
87
+ input_transforms: []
88
+ output_transforms: []
89
+ output_inv_transforms: []
90
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
91
+ noise_cfg: null
92
+ repo_id: horipse01/lerobot_merged_restricted_val
93
+ bbox_only: false
94
+ remove_bbox: false
95
+ episodes: null
96
+ splits:
97
+ - .*
98
+ eval_fraction: 0.1
99
+ eval_type: sample_mse
100
+ generation_train:
101
+ data:
102
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
103
+ root_path: null
104
+ action_horizon: null
105
+ state_history: null
106
+ input_transforms: []
107
+ output_transforms: []
108
+ output_inv_transforms: []
109
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
110
+ noise_cfg: null
111
+ repo_id: horipse01/lerobot_merged_restricted
112
+ bbox_only: true
113
+ remove_bbox: false
114
+ episodes: null
115
+ splits:
116
+ - .*
117
+ eval_fraction: 0.02
118
+ eval_type: generation
119
+ generation_test:
120
+ data:
121
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
122
+ root_path: null
123
+ action_horizon: null
124
+ state_history: null
125
+ input_transforms: []
126
+ output_transforms: []
127
+ output_inv_transforms: []
128
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
129
+ noise_cfg: null
130
+ repo_id: horipse01/lerobot_merged_restricted_val
131
+ bbox_only: true
132
+ remove_bbox: false
133
+ episodes: null
134
+ splits:
135
+ - .*
136
+ eval_fraction: 0.1
137
+ eval_type: generation
138
+ eval_num_sample_steps: 10
139
+ eval_batch_size: 32
140
+ policy:
141
+ _target_: vla_scratch.policies.pi.policy.PiPolicy
142
+ transforms:
143
+ - _target_: vla_scratch.policies.modules.vlm_bridge.qwen.processor.QwenProcessor
144
+ processor_class: Qwen3VLProcessor
145
+ model_id: Qwen/Qwen3-VL-2B-Instruct
146
+ max_length: 500
147
+ padding: max_length
148
+ state_history: 0
149
+ action_horizon: 10
150
+ state_dim: null
151
+ action_dim: null
152
+ vlm_type: Qwen3VLForConditionalGeneration
153
+ model_id: Qwen/Qwen3-VL-2B-Instruct
154
+ action_expert_cfg:
155
+ hidden_size: 1024
156
+ intermediate_size: 4096
157
+ num_attention_heads: 8
158
+ num_key_value_heads: 8
159
+ head_dim: 512
160
+ cross_attention_every: 2
161
+ qk_norm: layernorm
162
+ rotary_self_attn: true
163
+ only_attend_to_final_layer: true
164
+ attn_dropout: 0.0
165
+ mlp_dropout: 0.0
166
+ mlp_activation: silu
167
+ num_hidden_layers: 12
168
+ layers_for_dispersive_loss:
169
+ - 6
170
+ dispersive_loss_tau: 1.0
171
+ rms_norm_eps: 1.0e-06
172
+ attention_dropout: 0.0
173
+ attention_bias: true
174
+ max_position_embeddings: 8192
175
+ rope_theta: 10000.0
176
+ suffix_add_pos_emb: true
177
+ use_state: false
178
+ num_obs_registers: 4
179
+ expert_only_use_register: true
180
+ num_noise_per_sample: 2
181
+ num_noise_before_topk: 2
182
+ detach_kv_cache: false
183
+ ce_loss_weight: 0.1
184
+ disp_loss_weight: 0.0
185
+ time_dist_alpha: 1.0
186
+ time_dist_beta: 1.5
187
+ obs_register_init_gain: 0.02
188
+ suffix_pos_emb_init_gain: 0.02
189
+ zero_pos_id_for_obs_register: true
190
+ causal_mask_obs_register: true
191
+ qwen3_vl_use_grid_thw_list: true
192
+ qwen3_vl_recompute_pos_ids: false
193
+ qwen3_vl_masked_add_stack: true
194
+ checkpoint_path: null
195
+ load_optimizer: true
196
+ wandb:
197
+ project: vla-scratch
198
+ mode: online
199
+ tags: []
200
+ run_dir: /mnt/amlfs-01/home/hweng/projects/vla-scratch/vla-scratch/outputs/2025-12-25/19-55-39-pi-training
201
+ world_size: 8
train-cfg.yaml ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ defaults:
2
+ - _self_
3
+ - policy: pi-qwen
4
+ - data: libero-ipec
5
+ - train_data: none
6
+ - eval_data: none
7
+ num_workers: 4
8
+ prefetch_factor: 2
9
+ split_seed: 42
10
+ epoch_iterator: eager
11
+ epochs: 121
12
+ batch_size: 32
13
+ grad_accum_steps: 1
14
+ lr:
15
+ base: 5.0e-05
16
+ vlm_bridge: 1.0e-05
17
+ action_expert: 5.0e-05
18
+ warmup_steps: 0
19
+ cosine_anneal_epoch: 0
20
+ betas:
21
+ - 0.99
22
+ - 0.9999
23
+ eps: 1.0e-08
24
+ weight_decay: 0.0001
25
+ clip_grad_norm: 1.0
26
+ exp_name: pi-training
27
+ log_interval: 32
28
+ eval_interval: 512
29
+ save_interval: 40
30
+ data:
31
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
32
+ root_path: null
33
+ action_horizon: null
34
+ state_history: null
35
+ input_transforms: []
36
+ output_transforms: []
37
+ output_inv_transforms: []
38
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
39
+ - normalization_stats
40
+ - bbox_cotrain
41
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
42
+ noise_cfg: null
43
+ repo_id: horipse01/lerobot_merged_restricted
44
+ bbox_only: false
45
+ remove_bbox: false
46
+ episodes: null
47
+ splits:
48
+ - .*
49
+ train_data:
50
+ datasets:
51
+ action_a:
52
+ data:
53
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
54
+ root_path: null
55
+ action_horizon: 10
56
+ state_history: 0
57
+ input_transforms: []
58
+ output_transforms: []
59
+ output_inv_transforms: []
60
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
61
+ - normalization_stats
62
+ - bbox_cotrain
63
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
64
+ noise_cfg: null
65
+ repo_id: horipse01/lerobot_merged_restricted
66
+ bbox_only: false
67
+ remove_bbox: false
68
+ episodes: null
69
+ splits:
70
+ - .*
71
+ batch_size: 64
72
+ eval_data:
73
+ datasets:
74
+ action_train:
75
+ data:
76
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
77
+ root_path: null
78
+ action_horizon: 10
79
+ state_history: 0
80
+ input_transforms: []
81
+ output_transforms: []
82
+ output_inv_transforms: []
83
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
84
+ - normalization_stats
85
+ - bbox_cotrain
86
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
87
+ noise_cfg: null
88
+ repo_id: horipse01/lerobot_merged_restricted
89
+ bbox_only: false
90
+ remove_bbox: false
91
+ episodes: null
92
+ splits:
93
+ - .*
94
+ eval_fraction: 0.02
95
+ eval_type: sample_mse
96
+ action_test:
97
+ data:
98
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
99
+ root_path: null
100
+ action_horizon: 10
101
+ state_history: 0
102
+ input_transforms: []
103
+ output_transforms: []
104
+ output_inv_transforms: []
105
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
106
+ - normalization_stats
107
+ - bbox_cotrain
108
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
109
+ noise_cfg: null
110
+ repo_id: horipse01/lerobot_merged_restricted_val
111
+ bbox_only: false
112
+ remove_bbox: false
113
+ episodes: null
114
+ splits:
115
+ - .*
116
+ eval_fraction: 0.1
117
+ eval_type: sample_mse
118
+ generation_train:
119
+ data:
120
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
121
+ root_path: null
122
+ action_horizon: 10
123
+ state_history: 0
124
+ input_transforms: []
125
+ output_transforms: []
126
+ output_inv_transforms: []
127
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
128
+ - normalization_stats
129
+ - bbox_cotrain
130
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
131
+ noise_cfg: null
132
+ repo_id: horipse01/lerobot_merged_restricted
133
+ bbox_only: true
134
+ remove_bbox: false
135
+ episodes: null
136
+ splits:
137
+ - .*
138
+ eval_fraction: 0.02
139
+ eval_type: generation
140
+ generation_test:
141
+ data:
142
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
143
+ root_path: null
144
+ action_horizon: 10
145
+ state_history: 0
146
+ input_transforms: []
147
+ output_transforms: []
148
+ output_inv_transforms: []
149
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
150
+ - normalization_stats
151
+ - bbox_cotrain
152
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
153
+ noise_cfg: null
154
+ repo_id: horipse01/lerobot_merged_restricted_val
155
+ bbox_only: true
156
+ remove_bbox: false
157
+ episodes: null
158
+ splits:
159
+ - .*
160
+ eval_fraction: 0.1
161
+ eval_type: generation
162
+ eval_num_sample_steps: 10
163
+ eval_batch_size: 32
164
+ policy:
165
+ _target_: vla_scratch.policies.pi.policy.PiPolicy
166
+ transforms:
167
+ - _target_: vla_scratch.policies.modules.vlm_bridge.qwen.processor.QwenProcessor
168
+ processor_class: Qwen3VLProcessor
169
+ model_id: Qwen/Qwen3-VL-2B-Instruct
170
+ max_length: 500
171
+ padding: max_length
172
+ state_history: 0
173
+ action_horizon: 10
174
+ state_dim: 1
175
+ action_dim: 7
176
+ vlm_type: Qwen3VLForConditionalGeneration
177
+ model_id: Qwen/Qwen3-VL-2B-Instruct
178
+ action_expert_cfg:
179
+ hidden_size: 1024
180
+ intermediate_size: 4096
181
+ num_attention_heads: 8
182
+ num_key_value_heads: 8
183
+ head_dim: 512
184
+ cross_attention_every: 2
185
+ qk_norm: layernorm
186
+ rotary_self_attn: true
187
+ only_attend_to_final_layer: true
188
+ attn_dropout: 0.0
189
+ mlp_dropout: 0.0
190
+ mlp_activation: silu
191
+ num_hidden_layers: 12
192
+ layers_for_dispersive_loss:
193
+ - 6
194
+ dispersive_loss_tau: 1.0
195
+ rms_norm_eps: 1.0e-06
196
+ attention_dropout: 0.0
197
+ attention_bias: true
198
+ max_position_embeddings: 8192
199
+ rope_theta: 10000.0
200
+ suffix_add_pos_emb: true
201
+ use_state: false
202
+ num_obs_registers: 4
203
+ expert_only_use_register: true
204
+ num_noise_per_sample: 2
205
+ num_noise_before_topk: 2
206
+ detach_kv_cache: false
207
+ ce_loss_weight: 0.1
208
+ disp_loss_weight: 0.0
209
+ time_dist_alpha: 1.0
210
+ time_dist_beta: 1.5
211
+ obs_register_init_gain: 0.02
212
+ suffix_pos_emb_init_gain: 0.02
213
+ zero_pos_id_for_obs_register: true
214
+ causal_mask_obs_register: true
215
+ qwen3_vl_use_grid_thw_list: true
216
+ qwen3_vl_recompute_pos_ids: false
217
+ qwen3_vl_masked_add_stack: true
218
+ checkpoint_path: null
219
+ load_optimizer: true
220
+ wandb:
221
+ project: vla-scratch
222
+ mode: online
223
+ tags: []
224
+ hydra:
225
+ defaults:
226
+ - output: default
227
+ - launcher: basic
228
+ - sweeper: basic
229
+ - help: default
230
+ - hydra_help: default
231
+ - hydra_logging: default
232
+ - job_logging: default
233
+ - callbacks: null
234
+ - env: default
235
+ mode: null
236
+ searchpath: []
237
+ run:
238
+ dir: .
239
+ sweep:
240
+ dir: ???
241
+ subdir: ???
242
+ hydra_logging: ???
243
+ job_logging: ???
244
+ sweeper: ???
245
+ launcher: ???
246
+ callbacks: {}
247
+ help:
248
+ app_name: ???
249
+ header: ???
250
+ footer: ???
251
+ template: ???
252
+ hydra_help:
253
+ hydra_help: ???
254
+ template: ???
255
+ output_subdir: null
256
+ overrides:
257
+ hydra: []
258
+ task: []
259
+ job:
260
+ name: ???
261
+ chdir: false
262
+ override_dirname: ???
263
+ id: ???
264
+ num: ???
265
+ config_name: ???
266
+ env_set: {}
267
+ env_copy: []
268
+ config:
269
+ override_dirname:
270
+ kv_sep: '='
271
+ item_sep: ','
272
+ exclude_keys: []
273
+ runtime:
274
+ version: ???
275
+ version_base: ???
276
+ cwd: ???
277
+ config_sources: ???
278
+ output_dir: ???
279
+ choices: {}
280
+ verbose: false
wandb/run-20251225_195640-1lai13bo/files/cfg.yaml ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_workers: 4
2
+ prefetch_factor: 2
3
+ split_seed: 42
4
+ epoch_iterator: eager
5
+ epochs: 121
6
+ batch_size: 32
7
+ grad_accum_steps: 1
8
+ lr:
9
+ base: 5.0e-05
10
+ vlm_bridge: 1.0e-05
11
+ action_expert: 5.0e-05
12
+ warmup_steps: 0
13
+ cosine_anneal_epoch: 0
14
+ betas:
15
+ - 0.99
16
+ - 0.9999
17
+ eps: 1.0e-08
18
+ weight_decay: 0.0001
19
+ clip_grad_norm: 1.0
20
+ exp_name: pi-training
21
+ log_interval: 32
22
+ eval_interval: 512
23
+ save_interval: 40
24
+ data:
25
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
26
+ root_path: null
27
+ action_horizon: null
28
+ state_history: null
29
+ input_transforms: []
30
+ output_transforms: []
31
+ output_inv_transforms: []
32
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
33
+ noise_cfg: null
34
+ repo_id: horipse01/lerobot_merged_restricted
35
+ bbox_only: false
36
+ remove_bbox: false
37
+ episodes: null
38
+ splits:
39
+ - .*
40
+ train_data:
41
+ datasets:
42
+ action_a:
43
+ data:
44
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
45
+ root_path: null
46
+ action_horizon: null
47
+ state_history: null
48
+ input_transforms: []
49
+ output_transforms: []
50
+ output_inv_transforms: []
51
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
52
+ noise_cfg: null
53
+ repo_id: horipse01/lerobot_merged_restricted
54
+ bbox_only: false
55
+ remove_bbox: false
56
+ episodes: null
57
+ splits:
58
+ - .*
59
+ batch_size: 64
60
+ eval_data:
61
+ datasets:
62
+ action_train:
63
+ data:
64
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
65
+ root_path: null
66
+ action_horizon: null
67
+ state_history: null
68
+ input_transforms: []
69
+ output_transforms: []
70
+ output_inv_transforms: []
71
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
72
+ noise_cfg: null
73
+ repo_id: horipse01/lerobot_merged_restricted
74
+ bbox_only: false
75
+ remove_bbox: false
76
+ episodes: null
77
+ splits:
78
+ - .*
79
+ eval_fraction: 0.02
80
+ eval_type: sample_mse
81
+ action_test:
82
+ data:
83
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
84
+ root_path: null
85
+ action_horizon: null
86
+ state_history: null
87
+ input_transforms: []
88
+ output_transforms: []
89
+ output_inv_transforms: []
90
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
91
+ noise_cfg: null
92
+ repo_id: horipse01/lerobot_merged_restricted_val
93
+ bbox_only: false
94
+ remove_bbox: false
95
+ episodes: null
96
+ splits:
97
+ - .*
98
+ eval_fraction: 0.1
99
+ eval_type: sample_mse
100
+ generation_train:
101
+ data:
102
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
103
+ root_path: null
104
+ action_horizon: null
105
+ state_history: null
106
+ input_transforms: []
107
+ output_transforms: []
108
+ output_inv_transforms: []
109
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
110
+ noise_cfg: null
111
+ repo_id: horipse01/lerobot_merged_restricted
112
+ bbox_only: true
113
+ remove_bbox: false
114
+ episodes: null
115
+ splits:
116
+ - .*
117
+ eval_fraction: 0.02
118
+ eval_type: generation
119
+ generation_test:
120
+ data:
121
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
122
+ root_path: null
123
+ action_horizon: null
124
+ state_history: null
125
+ input_transforms: []
126
+ output_transforms: []
127
+ output_inv_transforms: []
128
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
129
+ noise_cfg: null
130
+ repo_id: horipse01/lerobot_merged_restricted_val
131
+ bbox_only: true
132
+ remove_bbox: false
133
+ episodes: null
134
+ splits:
135
+ - .*
136
+ eval_fraction: 0.1
137
+ eval_type: generation
138
+ eval_num_sample_steps: 10
139
+ eval_batch_size: 32
140
+ policy:
141
+ _target_: vla_scratch.policies.pi.policy.PiPolicy
142
+ transforms:
143
+ - _target_: vla_scratch.policies.modules.vlm_bridge.qwen.processor.QwenProcessor
144
+ processor_class: Qwen3VLProcessor
145
+ model_id: Qwen/Qwen3-VL-2B-Instruct
146
+ max_length: 500
147
+ padding: max_length
148
+ state_history: 0
149
+ action_horizon: 10
150
+ state_dim: null
151
+ action_dim: null
152
+ vlm_type: Qwen3VLForConditionalGeneration
153
+ model_id: Qwen/Qwen3-VL-2B-Instruct
154
+ action_expert_cfg:
155
+ hidden_size: 1024
156
+ intermediate_size: 4096
157
+ num_attention_heads: 8
158
+ num_key_value_heads: 8
159
+ head_dim: 512
160
+ cross_attention_every: 2
161
+ qk_norm: layernorm
162
+ rotary_self_attn: true
163
+ only_attend_to_final_layer: true
164
+ attn_dropout: 0.0
165
+ mlp_dropout: 0.0
166
+ mlp_activation: silu
167
+ num_hidden_layers: 12
168
+ layers_for_dispersive_loss:
169
+ - 6
170
+ dispersive_loss_tau: 1.0
171
+ rms_norm_eps: 1.0e-06
172
+ attention_dropout: 0.0
173
+ attention_bias: true
174
+ max_position_embeddings: 8192
175
+ rope_theta: 10000.0
176
+ suffix_add_pos_emb: true
177
+ use_state: false
178
+ num_obs_registers: 4
179
+ expert_only_use_register: true
180
+ num_noise_per_sample: 2
181
+ num_noise_before_topk: 2
182
+ detach_kv_cache: false
183
+ ce_loss_weight: 0.1
184
+ disp_loss_weight: 0.0
185
+ time_dist_alpha: 1.0
186
+ time_dist_beta: 1.5
187
+ obs_register_init_gain: 0.02
188
+ suffix_pos_emb_init_gain: 0.02
189
+ zero_pos_id_for_obs_register: true
190
+ causal_mask_obs_register: true
191
+ qwen3_vl_use_grid_thw_list: true
192
+ qwen3_vl_recompute_pos_ids: false
193
+ qwen3_vl_masked_add_stack: true
194
+ checkpoint_path: null
195
+ load_optimizer: true
196
+ wandb:
197
+ project: vla-scratch
198
+ mode: online
199
+ tags: []
200
+ run_dir: /mnt/amlfs-01/home/hweng/projects/vla-scratch/vla-scratch/outputs/2025-12-25/19-55-39-pi-training
201
+ world_size: 8
wandb/run-20251225_195640-1lai13bo/files/config.yaml ADDED
@@ -0,0 +1,344 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _wandb:
2
+ value:
3
+ cli_version: 0.21.4
4
+ e:
5
+ gt1vdj22y3nl3q7a422smxub7piq9xeu:
6
+ args:
7
+ - policy=pi-qwen
8
+ - policy.state_history=0
9
+ - policy.action_horizon=10
10
+ - policy.transforms.0.max_length=500
11
+ - data=bbox_cotrain_train
12
+ - batch_size=32
13
+ - train_data=bbox_cotrain_baseline
14
+ - train_data.datasets.action_a.batch_size=64
15
+ - eval_data=bbox_cotrain_eval
16
+ - num_workers=4
17
+ - prefetch_factor=2
18
+ - lr.base=5e-5
19
+ - +lr.vlm_bridge=1e-5
20
+ - +lr.action_expert=5e-5
21
+ - epochs=121
22
+ - save_interval=40
23
+ - wandb.mode=online
24
+ cpu_count: 96
25
+ cpu_count_logical: 96
26
+ cudaVersion: "13.0"
27
+ disk:
28
+ /:
29
+ total: "1065418129408"
30
+ used: "726712709120"
31
+ email: elijahgalahad@gmail.com
32
+ executable: /mnt/amlfs-01/home/hweng/projects/vla-scratch/vla-scratch/.venv/bin/python3
33
+ gpu: NVIDIA H100 80GB HBM3
34
+ gpu_count: 8
35
+ gpu_nvidia:
36
+ - architecture: Hopper
37
+ cudaCores: 16896
38
+ memoryTotal: "85520809984"
39
+ name: NVIDIA H100 80GB HBM3
40
+ uuid: GPU-83bb76f3-06ce-7f65-9aff-97f16dbac4ce
41
+ - architecture: Hopper
42
+ cudaCores: 16896
43
+ memoryTotal: "85520809984"
44
+ name: NVIDIA H100 80GB HBM3
45
+ uuid: GPU-6734b808-316d-13ad-3452-190c89899f8e
46
+ - architecture: Hopper
47
+ cudaCores: 16896
48
+ memoryTotal: "85520809984"
49
+ name: NVIDIA H100 80GB HBM3
50
+ uuid: GPU-bf6464b4-ee22-dd87-184c-579e33ac7b15
51
+ - architecture: Hopper
52
+ cudaCores: 16896
53
+ memoryTotal: "85520809984"
54
+ name: NVIDIA H100 80GB HBM3
55
+ uuid: GPU-d819de6f-6573-c624-1294-02a3acedf344
56
+ - architecture: Hopper
57
+ cudaCores: 16896
58
+ memoryTotal: "85520809984"
59
+ name: NVIDIA H100 80GB HBM3
60
+ uuid: GPU-88262183-ce36-a48f-dc17-3bde9f4de8f3
61
+ - architecture: Hopper
62
+ cudaCores: 16896
63
+ memoryTotal: "85520809984"
64
+ name: NVIDIA H100 80GB HBM3
65
+ uuid: GPU-98d3e6fa-f573-2b1a-79fd-01c89d027d23
66
+ - architecture: Hopper
67
+ cudaCores: 16896
68
+ memoryTotal: "85520809984"
69
+ name: NVIDIA H100 80GB HBM3
70
+ uuid: GPU-34bb43c6-303f-2a7c-4ede-99e51fc3e28f
71
+ - architecture: Hopper
72
+ cudaCores: 16896
73
+ memoryTotal: "85520809984"
74
+ name: NVIDIA H100 80GB HBM3
75
+ uuid: GPU-e30f59ec-7f5f-4c37-3417-33fccebb8823
76
+ host: 5c26df9320f14086-197be6b49fe441a3
77
+ memory:
78
+ total: "2008083251200"
79
+ os: Linux-5.15.0-1098-azure-x86_64-with-glibc2.35
80
+ program: /mnt/amlfs-01/home/hweng/projects/vla-scratch/vla-scratch/scripts/train_policy.py
81
+ python: CPython 3.10.12
82
+ root: /mnt/amlfs-01/home/hweng/projects/vla-scratch/vla-scratch/outputs/2025-12-25/19-55-39-pi-training
83
+ startedAt: "2025-12-26T03:56:40.589357Z"
84
+ writerId: gt1vdj22y3nl3q7a422smxub7piq9xeu
85
+ m: []
86
+ python_version: 3.10.12
87
+ t:
88
+ "1":
89
+ - 1
90
+ - 11
91
+ - 41
92
+ - 49
93
+ - 50
94
+ - 51
95
+ - 71
96
+ - 105
97
+ "2":
98
+ - 1
99
+ - 11
100
+ - 41
101
+ - 49
102
+ - 50
103
+ - 51
104
+ - 71
105
+ - 105
106
+ "3":
107
+ - 3
108
+ - 15
109
+ - 17
110
+ "4": 3.10.12
111
+ "5": 0.21.4
112
+ "6": 4.57.1
113
+ "12": 0.21.4
114
+ "13": linux-x86_64
115
+ batch_size:
116
+ value: 32
117
+ betas:
118
+ value:
119
+ - 0.99
120
+ - 0.9999
121
+ checkpoint_path:
122
+ value: null
123
+ clip_grad_norm:
124
+ value: 1
125
+ cosine_anneal_epoch:
126
+ value: 0
127
+ data:
128
+ value:
129
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
130
+ action_horizon: null
131
+ bbox_only: false
132
+ episodes: null
133
+ input_transforms: []
134
+ noise_cfg: null
135
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
136
+ output_inv_transforms: []
137
+ output_transforms: []
138
+ remove_bbox: false
139
+ repo_id: horipse01/lerobot_merged_restricted
140
+ root_path: null
141
+ splits:
142
+ - .*
143
+ state_history: null
144
+ epoch_iterator:
145
+ value: eager
146
+ epochs:
147
+ value: 121
148
+ eps:
149
+ value: 1e-08
150
+ eval_batch_size:
151
+ value: 32
152
+ eval_data:
153
+ value:
154
+ datasets:
155
+ action_test:
156
+ data:
157
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
158
+ action_horizon: null
159
+ bbox_only: false
160
+ episodes: null
161
+ input_transforms: []
162
+ noise_cfg: null
163
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
164
+ output_inv_transforms: []
165
+ output_transforms: []
166
+ remove_bbox: false
167
+ repo_id: horipse01/lerobot_merged_restricted_val
168
+ root_path: null
169
+ splits:
170
+ - .*
171
+ state_history: null
172
+ eval_fraction: 0.1
173
+ eval_type: sample_mse
174
+ action_train:
175
+ data:
176
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
177
+ action_horizon: null
178
+ bbox_only: false
179
+ episodes: null
180
+ input_transforms: []
181
+ noise_cfg: null
182
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
183
+ output_inv_transforms: []
184
+ output_transforms: []
185
+ remove_bbox: false
186
+ repo_id: horipse01/lerobot_merged_restricted
187
+ root_path: null
188
+ splits:
189
+ - .*
190
+ state_history: null
191
+ eval_fraction: 0.02
192
+ eval_type: sample_mse
193
+ generation_test:
194
+ data:
195
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
196
+ action_horizon: null
197
+ bbox_only: true
198
+ episodes: null
199
+ input_transforms: []
200
+ noise_cfg: null
201
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
202
+ output_inv_transforms: []
203
+ output_transforms: []
204
+ remove_bbox: false
205
+ repo_id: horipse01/lerobot_merged_restricted_val
206
+ root_path: null
207
+ splits:
208
+ - .*
209
+ state_history: null
210
+ eval_fraction: 0.1
211
+ eval_type: generation
212
+ generation_train:
213
+ data:
214
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
215
+ action_horizon: null
216
+ bbox_only: true
217
+ episodes: null
218
+ input_transforms: []
219
+ noise_cfg: null
220
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
221
+ output_inv_transforms: []
222
+ output_transforms: []
223
+ remove_bbox: false
224
+ repo_id: horipse01/lerobot_merged_restricted
225
+ root_path: null
226
+ splits:
227
+ - .*
228
+ state_history: null
229
+ eval_fraction: 0.02
230
+ eval_type: generation
231
+ eval_interval:
232
+ value: 512
233
+ eval_num_sample_steps:
234
+ value: 10
235
+ exp_name:
236
+ value: pi-training
237
+ grad_accum_steps:
238
+ value: 1
239
+ load_optimizer:
240
+ value: true
241
+ log_interval:
242
+ value: 32
243
+ lr:
244
+ value:
245
+ action_expert: 5e-05
246
+ base: 5e-05
247
+ vlm_bridge: 1e-05
248
+ num_workers:
249
+ value: 4
250
+ policy:
251
+ value:
252
+ _target_: vla_scratch.policies.pi.policy.PiPolicy
253
+ action_dim: null
254
+ action_expert_cfg:
255
+ attention_bias: true
256
+ attention_dropout: 0
257
+ attn_dropout: 0
258
+ cross_attention_every: 2
259
+ dispersive_loss_tau: 1
260
+ head_dim: 512
261
+ hidden_size: 1024
262
+ intermediate_size: 4096
263
+ layers_for_dispersive_loss:
264
+ - 6
265
+ max_position_embeddings: 8192
266
+ mlp_activation: silu
267
+ mlp_dropout: 0
268
+ num_attention_heads: 8
269
+ num_hidden_layers: 12
270
+ num_key_value_heads: 8
271
+ only_attend_to_final_layer: true
272
+ qk_norm: layernorm
273
+ rms_norm_eps: 1e-06
274
+ rope_theta: 10000
275
+ rotary_self_attn: true
276
+ action_horizon: 10
277
+ causal_mask_obs_register: true
278
+ ce_loss_weight: 0.1
279
+ detach_kv_cache: false
280
+ disp_loss_weight: 0
281
+ expert_only_use_register: true
282
+ model_id: Qwen/Qwen3-VL-2B-Instruct
283
+ num_noise_before_topk: 2
284
+ num_noise_per_sample: 2
285
+ num_obs_registers: 4
286
+ obs_register_init_gain: 0.02
287
+ qwen3_vl_masked_add_stack: true
288
+ qwen3_vl_recompute_pos_ids: false
289
+ qwen3_vl_use_grid_thw_list: true
290
+ state_dim: null
291
+ state_history: 0
292
+ suffix_add_pos_emb: true
293
+ suffix_pos_emb_init_gain: 0.02
294
+ time_dist_alpha: 1
295
+ time_dist_beta: 1.5
296
+ transforms:
297
+ - _target_: vla_scratch.policies.modules.vlm_bridge.qwen.processor.QwenProcessor
298
+ max_length: 500
299
+ model_id: Qwen/Qwen3-VL-2B-Instruct
300
+ padding: max_length
301
+ processor_class: Qwen3VLProcessor
302
+ use_state: false
303
+ vlm_type: Qwen3VLForConditionalGeneration
304
+ zero_pos_id_for_obs_register: true
305
+ prefetch_factor:
306
+ value: 2
307
+ run_dir:
308
+ value: /mnt/amlfs-01/home/hweng/projects/vla-scratch/vla-scratch/outputs/2025-12-25/19-55-39-pi-training
309
+ save_interval:
310
+ value: 40
311
+ split_seed:
312
+ value: 42
313
+ train_data:
314
+ value:
315
+ datasets:
316
+ action_a:
317
+ batch_size: 64
318
+ data:
319
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
320
+ action_horizon: null
321
+ bbox_only: false
322
+ episodes: null
323
+ input_transforms: []
324
+ noise_cfg: null
325
+ norm_stats_path: normalization_stats/bbox_cotrain/lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
326
+ output_inv_transforms: []
327
+ output_transforms: []
328
+ remove_bbox: false
329
+ repo_id: horipse01/lerobot_merged_restricted
330
+ root_path: null
331
+ splits:
332
+ - .*
333
+ state_history: null
334
+ wandb:
335
+ value:
336
+ mode: online
337
+ project: vla-scratch
338
+ tags: []
339
+ warmup_steps:
340
+ value: 0
341
+ weight_decay:
342
+ value: 0.0001
343
+ world_size:
344
+ value: 8
wandb/run-20251225_195640-1lai13bo/files/train-cfg.yaml ADDED
@@ -0,0 +1,280 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ defaults:
2
+ - _self_
3
+ - policy: pi-qwen
4
+ - data: libero-ipec
5
+ - train_data: none
6
+ - eval_data: none
7
+ num_workers: 4
8
+ prefetch_factor: 2
9
+ split_seed: 42
10
+ epoch_iterator: eager
11
+ epochs: 121
12
+ batch_size: 32
13
+ grad_accum_steps: 1
14
+ lr:
15
+ base: 5.0e-05
16
+ vlm_bridge: 1.0e-05
17
+ action_expert: 5.0e-05
18
+ warmup_steps: 0
19
+ cosine_anneal_epoch: 0
20
+ betas:
21
+ - 0.99
22
+ - 0.9999
23
+ eps: 1.0e-08
24
+ weight_decay: 0.0001
25
+ clip_grad_norm: 1.0
26
+ exp_name: pi-training
27
+ log_interval: 32
28
+ eval_interval: 512
29
+ save_interval: 40
30
+ data:
31
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
32
+ root_path: null
33
+ action_horizon: null
34
+ state_history: null
35
+ input_transforms: []
36
+ output_transforms: []
37
+ output_inv_transforms: []
38
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
39
+ - normalization_stats
40
+ - bbox_cotrain
41
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
42
+ noise_cfg: null
43
+ repo_id: horipse01/lerobot_merged_restricted
44
+ bbox_only: false
45
+ remove_bbox: false
46
+ episodes: null
47
+ splits:
48
+ - .*
49
+ train_data:
50
+ datasets:
51
+ action_a:
52
+ data:
53
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
54
+ root_path: null
55
+ action_horizon: 10
56
+ state_history: 0
57
+ input_transforms: []
58
+ output_transforms: []
59
+ output_inv_transforms: []
60
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
61
+ - normalization_stats
62
+ - bbox_cotrain
63
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
64
+ noise_cfg: null
65
+ repo_id: horipse01/lerobot_merged_restricted
66
+ bbox_only: false
67
+ remove_bbox: false
68
+ episodes: null
69
+ splits:
70
+ - .*
71
+ batch_size: 64
72
+ eval_data:
73
+ datasets:
74
+ action_train:
75
+ data:
76
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
77
+ root_path: null
78
+ action_horizon: 10
79
+ state_history: 0
80
+ input_transforms: []
81
+ output_transforms: []
82
+ output_inv_transforms: []
83
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
84
+ - normalization_stats
85
+ - bbox_cotrain
86
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
87
+ noise_cfg: null
88
+ repo_id: horipse01/lerobot_merged_restricted
89
+ bbox_only: false
90
+ remove_bbox: false
91
+ episodes: null
92
+ splits:
93
+ - .*
94
+ eval_fraction: 0.02
95
+ eval_type: sample_mse
96
+ action_test:
97
+ data:
98
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
99
+ root_path: null
100
+ action_horizon: 10
101
+ state_history: 0
102
+ input_transforms: []
103
+ output_transforms: []
104
+ output_inv_transforms: []
105
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
106
+ - normalization_stats
107
+ - bbox_cotrain
108
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
109
+ noise_cfg: null
110
+ repo_id: horipse01/lerobot_merged_restricted_val
111
+ bbox_only: false
112
+ remove_bbox: false
113
+ episodes: null
114
+ splits:
115
+ - .*
116
+ eval_fraction: 0.1
117
+ eval_type: sample_mse
118
+ generation_train:
119
+ data:
120
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
121
+ root_path: null
122
+ action_horizon: 10
123
+ state_history: 0
124
+ input_transforms: []
125
+ output_transforms: []
126
+ output_inv_transforms: []
127
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
128
+ - normalization_stats
129
+ - bbox_cotrain
130
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
131
+ noise_cfg: null
132
+ repo_id: horipse01/lerobot_merged_restricted
133
+ bbox_only: true
134
+ remove_bbox: false
135
+ episodes: null
136
+ splits:
137
+ - .*
138
+ eval_fraction: 0.02
139
+ eval_type: generation
140
+ generation_test:
141
+ data:
142
+ _target_: vla_scratch.datasets.bbox_cotrain.dataset.CoTrainDataset
143
+ root_path: null
144
+ action_horizon: 10
145
+ state_history: 0
146
+ input_transforms: []
147
+ output_transforms: []
148
+ output_inv_transforms: []
149
+ norm_stats_path: !!python/object/apply:pathlib.PosixPath
150
+ - normalization_stats
151
+ - bbox_cotrain
152
+ - lerobot_norm_stats-horizon_{data.action_horizon}-history_{data.state_history}.npz
153
+ noise_cfg: null
154
+ repo_id: horipse01/lerobot_merged_restricted_val
155
+ bbox_only: true
156
+ remove_bbox: false
157
+ episodes: null
158
+ splits:
159
+ - .*
160
+ eval_fraction: 0.1
161
+ eval_type: generation
162
+ eval_num_sample_steps: 10
163
+ eval_batch_size: 32
164
+ policy:
165
+ _target_: vla_scratch.policies.pi.policy.PiPolicy
166
+ transforms:
167
+ - _target_: vla_scratch.policies.modules.vlm_bridge.qwen.processor.QwenProcessor
168
+ processor_class: Qwen3VLProcessor
169
+ model_id: Qwen/Qwen3-VL-2B-Instruct
170
+ max_length: 500
171
+ padding: max_length
172
+ state_history: 0
173
+ action_horizon: 10
174
+ state_dim: 1
175
+ action_dim: 7
176
+ vlm_type: Qwen3VLForConditionalGeneration
177
+ model_id: Qwen/Qwen3-VL-2B-Instruct
178
+ action_expert_cfg:
179
+ hidden_size: 1024
180
+ intermediate_size: 4096
181
+ num_attention_heads: 8
182
+ num_key_value_heads: 8
183
+ head_dim: 512
184
+ cross_attention_every: 2
185
+ qk_norm: layernorm
186
+ rotary_self_attn: true
187
+ only_attend_to_final_layer: true
188
+ attn_dropout: 0.0
189
+ mlp_dropout: 0.0
190
+ mlp_activation: silu
191
+ num_hidden_layers: 12
192
+ layers_for_dispersive_loss:
193
+ - 6
194
+ dispersive_loss_tau: 1.0
195
+ rms_norm_eps: 1.0e-06
196
+ attention_dropout: 0.0
197
+ attention_bias: true
198
+ max_position_embeddings: 8192
199
+ rope_theta: 10000.0
200
+ suffix_add_pos_emb: true
201
+ use_state: false
202
+ num_obs_registers: 4
203
+ expert_only_use_register: true
204
+ num_noise_per_sample: 2
205
+ num_noise_before_topk: 2
206
+ detach_kv_cache: false
207
+ ce_loss_weight: 0.1
208
+ disp_loss_weight: 0.0
209
+ time_dist_alpha: 1.0
210
+ time_dist_beta: 1.5
211
+ obs_register_init_gain: 0.02
212
+ suffix_pos_emb_init_gain: 0.02
213
+ zero_pos_id_for_obs_register: true
214
+ causal_mask_obs_register: true
215
+ qwen3_vl_use_grid_thw_list: true
216
+ qwen3_vl_recompute_pos_ids: false
217
+ qwen3_vl_masked_add_stack: true
218
+ checkpoint_path: null
219
+ load_optimizer: true
220
+ wandb:
221
+ project: vla-scratch
222
+ mode: online
223
+ tags: []
224
+ hydra:
225
+ defaults:
226
+ - output: default
227
+ - launcher: basic
228
+ - sweeper: basic
229
+ - help: default
230
+ - hydra_help: default
231
+ - hydra_logging: default
232
+ - job_logging: default
233
+ - callbacks: null
234
+ - env: default
235
+ mode: null
236
+ searchpath: []
237
+ run:
238
+ dir: .
239
+ sweep:
240
+ dir: ???
241
+ subdir: ???
242
+ hydra_logging: ???
243
+ job_logging: ???
244
+ sweeper: ???
245
+ launcher: ???
246
+ callbacks: {}
247
+ help:
248
+ app_name: ???
249
+ header: ???
250
+ footer: ???
251
+ template: ???
252
+ hydra_help:
253
+ hydra_help: ???
254
+ template: ???
255
+ output_subdir: null
256
+ overrides:
257
+ hydra: []
258
+ task: []
259
+ job:
260
+ name: ???
261
+ chdir: false
262
+ override_dirname: ???
263
+ id: ???
264
+ num: ???
265
+ config_name: ???
266
+ env_set: {}
267
+ env_copy: []
268
+ config:
269
+ override_dirname:
270
+ kv_sep: '='
271
+ item_sep: ','
272
+ exclude_keys: []
273
+ runtime:
274
+ version: ???
275
+ version_base: ???
276
+ cwd: ???
277
+ config_sources: ???
278
+ output_dir: ???
279
+ choices: {}
280
+ verbose: false