outsider86 commited on
Commit
f0527fe
·
verified ·
1 Parent(s): a4b468e

Upload starVLA checkpoint classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k step 30000

Browse files
StarVLA/classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k/checkpoints/steps_30000_pytorch_model.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4c6a948e9df26ff8b5de68e93a92b27a0a5985ee7570836c6ff0fa0a89781050
3
+ size 5848235293
StarVLA/classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k/config.full.yaml ADDED
@@ -0,0 +1,133 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ run_id: classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k
2
+ run_root_dir: /scratch/wangpc/DiscreteRTCv2/results/Checkpoints
3
+ seed: 42
4
+ wandb_entity: 401910710-university-of-california-berkeley
5
+ wandb_project: starVLA_piper
6
+ is_debug: false
7
+ version_id: '0.21'
8
+ framework:
9
+ name: QwenDiscreteDiffusion
10
+ state_conditioning: instruction
11
+ action_tokens:
12
+ vocab_size: 512
13
+ bounds_file: ./data/processed/3cam/StateDeltaJoint/classify_blocks_3cam_30hz_cleanup_quantized_v512/quantizer.json
14
+ low:
15
+ - -0.055921837634838094
16
+ - -0.10420534206974891
17
+ - -0.09092266779826827
18
+ - -0.059009552001953125
19
+ - -0.07047413613088999
20
+ - -0.06176897424917955
21
+ - 0.004000000189989805
22
+ high:
23
+ - 0.0694117546081543
24
+ - 0.09826207160949707
25
+ - 0.0864112377166748
26
+ - 0.07096407331269362
27
+ - 0.12524479627609253
28
+ - 0.06398379802703857
29
+ - 1.0
30
+ self_distillation: null
31
+ qwenvl:
32
+ base_vlm: ./playground/Pretrained_models/Qwen3-VL-2B-Instruct-Action
33
+ attn_implementation: flash_attention_2
34
+ vl_hidden_dim: 2048
35
+ num_vl_layers: 28
36
+ action_model:
37
+ action_model_type: LayerwiseDD
38
+ action_dim: 7
39
+ state_dim: 7
40
+ action_horizon: 30
41
+ repeated_diffusion_steps: 2
42
+ add_pos_embed: true
43
+ max_seq_len: 1024
44
+ num_target_vision_tokens: 32
45
+ representation: bin
46
+ num_bins: 512
47
+ action_low: -1.0
48
+ action_high: 1.0
49
+ num_inference_steps: 8
50
+ train_mask_schedule: cosine
51
+ decode_schedule: cosine
52
+ no_mask_token_prob: 0.0
53
+ l1_loss_weight: 0.0
54
+ use_simple_max: false
55
+ target_space: tokens
56
+ block_diffusion: null
57
+ diffusion_model_cfg:
58
+ action_dit_hidden_dim: 1024
59
+ dropout: 0.2
60
+ final_dropout: true
61
+ interleave_self_attention: true
62
+ norm_type: ada_norm
63
+ positional_embeddings: null
64
+ attention_head_dim: 64
65
+ num_layers: 28
66
+ input_embedding_dim: 1024
67
+ cross_attention_dim: 1024
68
+ num_attention_heads: 16
69
+ output_dim: 512
70
+ future_action_window_size: 29
71
+ past_action_window_size: 0
72
+ datasets:
73
+ vla_data:
74
+ dataset_py: lerobot_datasets
75
+ data_root_dir: ./data
76
+ data_mix: classify_blocks_3cam_30hz_delta_joint_raw_h30_v512
77
+ tokens_dataset_path: ./data/processed/3cam/StateDeltaJoint/classify_blocks_3cam_30hz_cleanup_quantized_v512
78
+ include_state: true
79
+ episode_split: train
80
+ action_type: delta_qpos
81
+ action_mode: abs
82
+ sequential_step_sampling: false
83
+ per_device_batch_size: 16
84
+ load_all_data_for_training: true
85
+ obs_image_size:
86
+ - 224
87
+ - 224
88
+ delete_pause_frame: false
89
+ video_backend: torchvision_av
90
+ num_workers: 2
91
+ persistent_workers: true
92
+ prefetch_factor: 2
93
+ pin_memory: true
94
+ drop_incomplete_tail_windows: 30
95
+ trainer:
96
+ pretrained_checkpoint: null
97
+ freeze_modules: qwen_vl_interface
98
+ max_train_steps: 30000
99
+ num_warmup_steps: 500
100
+ save_interval: 10000
101
+ eval_interval: 1000
102
+ repeated_diffusion_steps: 2
103
+ learning_rate:
104
+ base: 1.0e-05
105
+ qwen_vl_interface: 1.0e-06
106
+ action_model: 0.0001
107
+ project_layers: 0.0001
108
+ lr_scheduler_type: cosine_with_min_lr
109
+ scheduler_specific_kwargs:
110
+ min_lr: 1.0e-06
111
+ loss_scale:
112
+ vla: 1.0
113
+ vlm: 0.0
114
+ max_grad_norm: 1.0
115
+ weight_decay: 0.0
116
+ logging_frequency: 20
117
+ gradient_clipping: 1.0
118
+ gradient_accumulation_steps: 1
119
+ gradient_checkpointing: true
120
+ optimizer:
121
+ name: AdamW
122
+ betas:
123
+ - 0.9
124
+ - 0.95
125
+ eps: 1.0e-08
126
+ weight_decay: 1.0e-08
127
+ save_format: pt
128
+ is_resume: false
129
+ max_wall_time_hours: 48
130
+ wall_time_safety_minutes: 20
131
+ save_training_state: true
132
+ config_yaml: /scratch/wangpc/DiscreteRTCv2/examples/realRobots/Piper/train_files/starvla_qwendd_qwen3vl2b_classify_blocks_3cam_30hz_delta_joint_raw_h30_v512_2gpu_gb32_30k.yaml
133
+ output_dir: /scratch/wangpc/DiscreteRTCv2/results/Checkpoints/classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k
StarVLA/classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k/config.yaml ADDED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ datasets:
2
+ vla_data:
3
+ action_mode: abs
4
+ data_mix: classify_blocks_3cam_30hz_delta_joint_raw_h30_v512
5
+ data_root_dir: ./data
6
+ dataset_py: lerobot_datasets
7
+ delete_pause_frame: false
8
+ drop_incomplete_tail_windows: 30
9
+ episode_split: train
10
+ num_workers: 2
11
+ obs_image_size:
12
+ - 224
13
+ - 224
14
+ per_device_batch_size: 16
15
+ persistent_workers: true
16
+ pin_memory: true
17
+ prefetch_factor: 2
18
+ tokens_dataset_path: ./data/processed/3cam/StateDeltaJoint/classify_blocks_3cam_30hz_cleanup_quantized_v512
19
+ video_backend: torchvision_av
20
+ framework:
21
+ action_model:
22
+ action_dim: 7
23
+ action_high: 1.0
24
+ action_horizon: 30
25
+ action_low: -1.0
26
+ add_pos_embed: true
27
+ block_diffusion: null
28
+ decode_schedule: cosine
29
+ diffusion_model_cfg:
30
+ action_dit_hidden_dim: 1024
31
+ attention_head_dim: 64
32
+ cross_attention_dim: 1024
33
+ dropout: 0.2
34
+ final_dropout: true
35
+ input_embedding_dim: 1024
36
+ interleave_self_attention: true
37
+ norm_type: ada_norm
38
+ num_attention_heads: 16
39
+ num_layers: 28
40
+ output_dim: 512
41
+ positional_embeddings: null
42
+ l1_loss_weight: 0.0
43
+ max_seq_len: 1024
44
+ no_mask_token_prob: 0.0
45
+ num_bins: 512
46
+ num_inference_steps: 8
47
+ num_target_vision_tokens: 32
48
+ repeated_diffusion_steps: 2
49
+ representation: bin
50
+ state_dim: 7
51
+ target_space: tokens
52
+ train_mask_schedule: cosine
53
+ use_simple_max: false
54
+ action_tokens:
55
+ bounds_file: ./data/processed/3cam/StateDeltaJoint/classify_blocks_3cam_30hz_cleanup_quantized_v512/quantizer.json
56
+ high:
57
+ - 0.0694117546081543
58
+ - 0.09826207160949707
59
+ - 0.0864112377166748
60
+ - 0.07096407331269362
61
+ - 0.12524479627609253
62
+ - 0.06398379802703857
63
+ - 1.0
64
+ low:
65
+ - -0.055921837634838094
66
+ - -0.10420534206974891
67
+ - -0.09092266779826827
68
+ - -0.059009552001953125
69
+ - -0.07047413613088999
70
+ - -0.06176897424917955
71
+ - 0.004000000189989805
72
+ vocab_size: 512
73
+ name: QwenDiscreteDiffusion
74
+ qwenvl:
75
+ attn_implementation: flash_attention_2
76
+ base_vlm: ./playground/Pretrained_models/Qwen3-VL-2B-Instruct-Action
77
+ num_vl_layers: 28
78
+ vl_hidden_dim: 2048
79
+ self_distillation: null
80
+ state_conditioning: instruction
81
+ output_dir: /scratch/wangpc/DiscreteRTCv2/results/Checkpoints/classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k
82
+ run_id: classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k
83
+ run_root_dir: /scratch/wangpc/DiscreteRTCv2/results/Checkpoints
84
+ seed: 42
85
+ trainer:
86
+ eval_interval: 1000
87
+ freeze_modules: qwen_vl_interface
88
+ gradient_clipping: 1.0
89
+ is_resume: false
90
+ learning_rate:
91
+ action_model: 0.0001
92
+ base: 1.0e-05
93
+ project_layers: 0.0001
94
+ qwen_vl_interface: 1.0e-06
95
+ logging_frequency: 20
96
+ lr_scheduler_type: cosine_with_min_lr
97
+ max_train_steps: 30000
98
+ max_wall_time_hours: 48
99
+ num_warmup_steps: 500
100
+ optimizer:
101
+ betas:
102
+ - 0.9
103
+ - 0.95
104
+ eps: 1.0e-08
105
+ weight_decay: 1.0e-08
106
+ pretrained_checkpoint: null
107
+ save_format: pt
108
+ save_interval: 10000
109
+ save_training_state: true
110
+ scheduler_specific_kwargs:
111
+ min_lr: 1.0e-06
112
+ wall_time_safety_minutes: 20
113
+ wandb_entity: 401910710-university-of-california-berkeley
114
+ wandb_project: starVLA_piper
StarVLA/classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k/dataset_statistics.json ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "new_embodiment": {
3
+ "action": {
4
+ "mean": [
5
+ -2.5492838631180348e-06,
6
+ 1.8818147509591654e-05,
7
+ -2.4244372980319895e-05,
8
+ 9.33146002353169e-05,
9
+ 3.238847784814425e-05,
10
+ -0.00010508225386729464,
11
+ 0.3567821681499481
12
+ ],
13
+ "std": [
14
+ 0.007420769426971674,
15
+ 0.01777242124080658,
16
+ 0.015129923820495605,
17
+ 0.007751822005957365,
18
+ 0.010863296687602997,
19
+ 0.005989530589431524,
20
+ 0.33469298481941223
21
+ ],
22
+ "max": [
23
+ 0.0694117546081543,
24
+ 0.09826207160949707,
25
+ 0.0864112377166748,
26
+ 0.07079055905342102,
27
+ 0.12524479627609253,
28
+ 0.06398379802703857,
29
+ 1.0
30
+ ],
31
+ "min": [
32
+ -0.05579817667603493,
33
+ -0.10403919219970703,
34
+ -0.0906524658203125,
35
+ -0.059009552001953125,
36
+ -0.0702669620513916,
37
+ -0.06157521530985832,
38
+ 0.004000000189989805
39
+ ],
40
+ "q01": [
41
+ -0.023735604435205455,
42
+ -0.044436097145080566,
43
+ -0.036180734634399414,
44
+ -0.021013736724853516,
45
+ -0.03036874532699585,
46
+ -0.01963496208190918,
47
+ 0.004000000189989805
48
+ ],
49
+ "q99": [
50
+ 0.022165685892105103,
51
+ 0.03803074359893799,
52
+ 0.042359113693237305,
53
+ 0.027663469314575195,
54
+ 0.030333814024925233,
55
+ 0.017994344234466553,
56
+ 1.0
57
+ ],
58
+ "mask": [
59
+ true,
60
+ true,
61
+ true,
62
+ true,
63
+ true,
64
+ true,
65
+ true
66
+ ]
67
+ },
68
+ "state": {
69
+ "mean": [
70
+ -0.15684281289577484,
71
+ 1.7120212316513062,
72
+ -1.3335168361663818,
73
+ 0.24333560466766357,
74
+ 0.6909483671188354,
75
+ -0.5095524787902832,
76
+ 0.35678568482398987
77
+ ],
78
+ "std": [
79
+ 0.3099547028541565,
80
+ 0.5250895619392395,
81
+ 0.518146276473999,
82
+ 0.30827057361602783,
83
+ 0.33267882466316223,
84
+ 0.41217276453971863,
85
+ 0.334687739610672
86
+ ],
87
+ "max": [
88
+ 0.5265658497810364,
89
+ 2.7064297199249268,
90
+ -0.0005585053586401045,
91
+ 1.51981520652771,
92
+ 1.2243659496307373,
93
+ 1.0816328525543213,
94
+ 1.0
95
+ ],
96
+ "min": [
97
+ -0.9799150824546814,
98
+ 0.0009773843921720982,
99
+ -2.854311466217041,
100
+ -0.6770481467247009,
101
+ -0.4627915024757385,
102
+ -1.874780297279358,
103
+ 0.004000000189989805
104
+ ],
105
+ "q01": [
106
+ -0.7300013899803162,
107
+ 0.06363470107316971,
108
+ -2.499322748184204,
109
+ -0.341019868850708,
110
+ -0.11265838593244552,
111
+ -1.491436243057251,
112
+ 0.004000000189989805
113
+ ],
114
+ "q99": [
115
+ 0.36828018575906796,
116
+ 2.4737590193748478,
117
+ -0.008429939858615398,
118
+ 1.1437840461730957,
119
+ 1.1804525494575502,
120
+ 0.28834322541952195,
121
+ 1.0
122
+ ]
123
+ },
124
+ "num_transitions": 204691,
125
+ "num_trajectories": 87
126
+ }
127
+ }
StarVLA/classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k/summary.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {"steps": 10000}
2
+ {"steps": 20000}
3
+ {"steps": 30000}
StarVLA/classify_blocks_3cam_30hz_qwendd_qwen3vl2b_delta_joint_raw_h30_v512_2gpu_gb32_30k/wall_time_ledger.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "budget_hours": 48.0,
3
+ "safety_minutes": 20.0,
4
+ "sessions": [
5
+ {
6
+ "started_at": "2026-11-06T10:27:10+00:00",
7
+ "elapsed_seconds": 47168.6,
8
+ "start_step": 0,
9
+ "last_step": 30000,
10
+ "end_reason": "completed",
11
+ "signal": null
12
+ }
13
+ ],
14
+ "total_elapsed_seconds": 47168.6
15
+ }