TingtingDu commited on
Commit
8216272
·
verified ·
1 Parent(s): 127e9cb

Upload folder using huggingface_hub

Browse files
beta/metadata.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:980b85d74b7a4f32d5227b3890e3317f750b9f1c7b255162ac9ed9a72734f88c
3
+ size 1267
beta/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d7c162b14275476cf9cf050c8a88527db1c28bbaa1740aa9b1a307e470ff0bdf
3
+ size 7874279060
beta/norm_stats.json ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "state": {
3
+ "mean": [
4
+ -0.046521167204962556,
5
+ 0.03440711916118233,
6
+ 0.7645510541638811,
7
+ 2.972206629083346,
8
+ -0.22048456658605764,
9
+ -0.12558510205548704,
10
+ 0.026914437835321277,
11
+ -0.027190919539383258
12
+ ],
13
+ "std": [
14
+ 0.10494495178849877,
15
+ 0.15176981509729204,
16
+ 0.3785169899678948,
17
+ 0.34427321172541686,
18
+ 0.9069665667914908,
19
+ 0.3253953728031856,
20
+ 0.014175856715986892,
21
+ 0.014058836306644338
22
+ ],
23
+ "q01": [
24
+ -0.3993664956152439,
25
+ -0.2691648515164852,
26
+ 0.0380015922319144,
27
+ 1.5083277968406676,
28
+ -2.7224448468208315,
29
+ -1.0813218128442763,
30
+ 0.00173520012809895,
31
+ -0.04005292297729757
32
+ ],
33
+ "q99": [
34
+ 0.13545873028039934,
35
+ 0.33551978799700743,
36
+ 1.27041562808156,
37
+ 3.2765059180259706,
38
+ 2.4039963840961454,
39
+ 0.5971528978824616,
40
+ 0.04030334069116507,
41
+ -0.0017878379381960213
42
+ ]
43
+ },
44
+ "action": {
45
+ "mean": [
46
+ 0.05515142225915216,
47
+ 0.08769114746846801,
48
+ -0.08734487482305489,
49
+ 0.0007091030675075027,
50
+ 0.005476403673860704,
51
+ -0.005540392002278206,
52
+ -0.04563926272674084
53
+ ],
54
+ "std": [
55
+ 0.33186087749168136,
56
+ 0.37579685333032686,
57
+ 0.44811232017091995,
58
+ 0.03933290297384835,
59
+ 0.0635858615556415,
60
+ 0.0785736716640555,
61
+ 0.998957985952242
62
+ ],
63
+ "q01": [
64
+ -0.7072499999999999,
65
+ -0.7983750000000001,
66
+ -0.9375,
67
+ -0.11467779144644738,
68
+ -0.16395,
69
+ -0.2240490057706833,
70
+ -1.0
71
+ ],
72
+ "q99": [
73
+ 0.937125,
74
+ 0.86775,
75
+ 0.937125,
76
+ 0.13175314307212832,
77
+ 0.19169999999999998,
78
+ 0.33742949963808055,
79
+ 0.9996
80
+ ]
81
+ }
82
+ }
beta/optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5d179e5eca17513ddc03dba6628f489498219f3da505017f63f7bfc1899d30b9
3
+ size 2842889522
beta/train_config.yaml ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ runtime:
2
+ exp_name: beta-0330
3
+ checkpoint_base_dir: ./checkpoints
4
+ wandb_project: beta-0330
5
+ wandb_enabled: true
6
+ seed: 42
7
+ batch_size: 16
8
+ grad_accumulation_steps: 1
9
+ use_amp: true
10
+ num_train_steps: 100000
11
+ log_interval: 20
12
+ save_interval: 10000
13
+ learning_rate: 0.0002
14
+ weight_decay: 0.0001
15
+ num_workers: 4
16
+ resume: true
17
+ resume_from_best: false
18
+ overwrite: false
19
+ grad_clip_norm: 1.0
20
+ warmup_steps: 500
21
+ lr_schedule: cosine_warmup
22
+ end_lr: 1.0e-06
23
+ data:
24
+ repo_id: physical-intelligence/libero
25
+ split: train
26
+ num_workers: 4
27
+ max_token_len: 128
28
+ max_samples: null
29
+ state_dim: 8
30
+ norm_stats_path: assets/physical-intelligence/libero/norm_stats.json
31
+ temporal_frames: 1
32
+ temporal_stride: 5
33
+ model:
34
+ action_dim: 7
35
+ action_horizon: 10
36
+ state_dim: 8
37
+ gripper_loss_weight: 1.0
38
+ freeze_vision: false
39
+ freeze_language: false
40
+ freeze_vggt: false
41
+ use_lora: true
42
+ lora_r: 16
43
+ lora_alpha: 32
44
+ lora_dropout: 0.05
45
+ lora_target_modules:
46
+ - q_proj
47
+ - k_proj
48
+ - v_proj
49
+ - o_proj
50
+ - up_proj
51
+ - down_proj
52
+ - gate_proj
53
+ lora_target_modules_vggt:
54
+ - qkv
55
+ - proj
56
+ - fc1
57
+ - fc2
58
+ lora_on_language: true
59
+ lora_on_vggt: true
60
+ lora_on_vision: true
61
+ lora_target_modules_vision:
62
+ - q_proj
63
+ - k_proj
64
+ - v_proj
65
+ - out_proj
66
+ - fc1
67
+ - fc2
68
+ vision:
69
+ model_name: google/paligemma2-3b-pt-224
70
+ image_size: 224
71
+ language:
72
+ model_name: Qwen/Qwen3-0.6B-Base
73
+ trust_remote_code: true
74
+ vggt:
75
+ model_name: facebook/VGGT-1B
76
+ trust_remote_code: true
77
+ vision_patch_hw: !!python/tuple
78
+ - 16
79
+ - 16
80
+ num_cameras: 2
81
+ temporal_frames: 1