Evanxw0715 commited on
Commit
2cbd0bf
·
verified ·
1 Parent(s): 1cfb15f

Upload BOAT/BOAT_tokenizer

Browse files
BOAT/BOAT_tokenizer/README.md ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - model_hub_mixin
4
+ - pytorch_model_hub_mixin
5
+ ---
6
+
7
+ This model has been pushed to the Hub using the [PytorchModelHubMixin](https://huggingface.co/docs/huggingface_hub/package_reference/mixins#huggingface_hub.PyTorchModelHubMixin) integration:
8
+ - Code: [More Information Needed]
9
+ - Paper: [More Information Needed]
10
+ - Docs: [More Information Needed]
BOAT/BOAT_tokenizer/config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "action_dim": 14,
3
+ "action_horizon": 32,
4
+ "decoder_depth": 6,
5
+ "dropout": 0.1,
6
+ "emb_dim": 256,
7
+ "encoder_depth": 6,
8
+ "head_dim": 32,
9
+ "latent_dim": 4,
10
+ "latent_dropout_mode": "pow2",
11
+ "latent_horizon": 16,
12
+ "name": "boat",
13
+ "pretrained_path": null,
14
+ "quantizer_apply_corrupt_tokens_p": 0.2,
15
+ "quantizer_corrupt_tokens_p": 0.0,
16
+ "quantizer_drop_quant_p": 0.0,
17
+ "quantizer_levels": [
18
+ 8,
19
+ 8,
20
+ 6,
21
+ 5
22
+ ],
23
+ "quantizer_min_corrupt_tokens_p": null
24
+ }
BOAT/BOAT_tokenizer/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:76a842029f63a85802d63082d27b2dd98ffaf75074247e3fb20bbe932ad374a9
3
+ size 44181600
BOAT/BOAT_tokenizer/norm_stats.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4b9e7e69a4cb1bf67acc655cba23f329c1a6125d8c74ee1b37c19bd2c54a56e5
3
+ size 3957
BOAT/BOAT_tokenizer/policy_io.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "obs_features": [],
3
+ "action_feature": {
4
+ "key": "action",
5
+ "type": "action",
6
+ "shape": [
7
+ 14
8
+ ],
9
+ "norm_mode": "min_max"
10
+ },
11
+ "n_obs_steps": 0,
12
+ "n_action_steps": 32,
13
+ "fps": 15,
14
+ "language_key": null
15
+ }
BOAT/BOAT_tokenizer/train_config.yaml ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ codec:
2
+ name: boat
3
+ kind: tokens
4
+ pretrained_path: null
5
+ action_dim: null
6
+ action_horizon: 32
7
+ max_fit_chunks: 50000
8
+ latent_horizon: 16
9
+ emb_dim: 256
10
+ head_dim: 32
11
+ encoder_depth: 6
12
+ decoder_depth: 6
13
+ dropout: 0.1
14
+ latent_dropout_mode: pow2
15
+ quantizer_levels:
16
+ - 8
17
+ - 8
18
+ - 6
19
+ - 5
20
+ quantizer_drop_quant_p: 0.0
21
+ quantizer_corrupt_tokens_p: 0.0
22
+ quantizer_min_corrupt_tokens_p: null
23
+ quantizer_apply_corrupt_tokens_p: 0.2
24
+ dataset:
25
+ name: robotwin_new_demo_clean
26
+ repo_id: local/robotwin_new_demo_clean
27
+ root: /scratch/xuanweiliu/bimanual_tokenizer/data/lerobot/local/robotwin_new_demo_clean
28
+ hf_datasets_cache: ${oc.env:HF_DATASETS_CACHE,null}
29
+ episodes: null
30
+ exclude_episodes_file: null
31
+ action_stats_file: null
32
+ observation_keys:
33
+ - observation.state
34
+ - observation.images.cam_high
35
+ - observation.images.cam_left_wrist
36
+ - observation.images.cam_right_wrist
37
+ video_backend: null
38
+ bypass_force_cache_sync: true
39
+ trainer:
40
+ max_steps: 300000
41
+ batch_size: 512
42
+ global_batch_size: null
43
+ num_workers: 4
44
+ prefetch_factor: 4
45
+ use_episode_aware_sampler: true
46
+ drop_n_first_frames: 0
47
+ drop_n_last_frames: 0
48
+ weight_decay: 0.0
49
+ betas:
50
+ - 0.9
51
+ - 0.95
52
+ eps: 1.0e-08
53
+ lr_scheduler: constant
54
+ warmup_steps: 200
55
+ min_lr_ratio: 0.1
56
+ use_ema: true
57
+ ema_decay: 0.995
58
+ mixed_precision: bf16
59
+ gradient_accumulation_steps: 1
60
+ find_unused_parameters: false
61
+ max_grad_norm: 1.0
62
+ distributed_strategy: ddp
63
+ fsdp_min_num_params: 100000000
64
+ fsdp_reshard_after_forward: true
65
+ fsdp_cpu_offload: false
66
+ log_every: 200
67
+ val_every: 1000
68
+ checkpoint_every: 5000
69
+ checkpoint_dir: ${output_dir}/checkpoints
70
+ local_log_path: ${output_dir}/train_metrics.jsonl
71
+ val_ratio: 0.05
72
+ split_seed: ${seed}
73
+ val_batch_size: null
74
+ max_val_batches: 100
75
+ checkpoint_candidates:
76
+ best:
77
+ metric_key: val_reconst_loss
78
+ mode: min
79
+ k: 3
80
+ recent:
81
+ k: 2
82
+ use_wandb: true
83
+ wandb_project: tokenizer_pretrain
84
+ wandb_entity: null
85
+ wandb_init_timeout: null
86
+ wandb_upload_eval_videos: true
87
+ wandb_max_eval_videos: 16
88
+ _trainer_type: codecs
89
+ lr: 5.0e-05
90
+ reconst_every: 1000
91
+ max_reconst_batches: 100
92
+ skip_codec_fit: false
93
+ name: train_codecs
94
+ seed: 42
95
+ device: cuda
96
+ run_name: robotwin_new_boat_tokenizer
97
+ output_dir: /scratch/xuanweiliu/bimanual_tokenizer/outputs/hopper_boat_robotwin_new/616167.hopper-m-02/train
98
+ run_mode: overwrite
99
+ resume_checkpoint: null