codemanCheng commited on
Commit
c68c1fc
·
verified ·
1 Parent(s): 290c5c5

Add asset README

Browse files
Files changed (1) hide show
  1. README.md +166 -0
README.md ADDED
@@ -0,0 +1,166 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # PhysicalWord last05_mot2_action Assets
2
+
3
+ This dataset repo contains the private/custom assets needed to train
4
+ `last05_mot2_action` on a machine that cannot access the source NAS.
5
+
6
+ Code repo expected by the migration guide:
7
+
8
+ - GitHub: `https://github.com/zhang-yi-ming/phyword.git`
9
+ - Branch: `mot2_cosmos_ae`
10
+ - Commit: `1495a9d`
11
+
12
+ ## Included
13
+
14
+ This repo intentionally includes:
15
+
16
+ ```text
17
+ ckpt/pretrained/LaST0_Pretrain_AE_chunk16/tfmr/
18
+ data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual/
19
+ rlbench/
20
+ ```
21
+
22
+ Approximate source sizes:
23
+
24
+ ```text
25
+ ckpt/pretrained/LaST0_Pretrain_AE_chunk16/tfmr 4.0G
26
+ data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual 2.8G
27
+ rlbench 25G
28
+ ```
29
+
30
+ ## Not Included
31
+
32
+ Cosmos Predict2.5 weights are not included in this repo. Download them from
33
+ the official gated NVIDIA Hugging Face model after accepting the license:
34
+
35
+ ```bash
36
+ huggingface-cli download nvidia/Cosmos-Predict2.5-2B \
37
+ --local-dir /mnt/nas/zhangyiming/database/ckpt/pretrained/Cosmos-Predict2.5-2B
38
+ ```
39
+
40
+ The source scripts expect the Cosmos checkpoint at:
41
+
42
+ ```text
43
+ /mnt/nas/zhangyiming/database/ckpt/pretrained/Cosmos-Predict2.5-2B/base/pre-trained/d20b7120-df3e-4911-919d-db6e08bad31c_ema_bf16.pt
44
+ ```
45
+
46
+ Already-trained MoT2 evaluation checkpoints are also not included.
47
+
48
+ ## Target Layout
49
+
50
+ Download this dataset repo directly into:
51
+
52
+ ```bash
53
+ /mnt/nas/zhangyiming/database
54
+ ```
55
+
56
+ Example:
57
+
58
+ ```bash
59
+ export HF_HUB_ENABLE_HF_TRANSFER=1
60
+ mkdir -p /mnt/nas/zhangyiming/database
61
+
62
+ huggingface-cli download <owner>/physicalword-assets \
63
+ --repo-type dataset \
64
+ --local-dir /mnt/nas/zhangyiming/database
65
+ ```
66
+
67
+ The training JSON files contain absolute paths under
68
+ `/mnt/nas/zhangyiming/database`. If the assets are downloaded elsewhere, create
69
+ a symlink:
70
+
71
+ ```bash
72
+ mkdir -p /mnt/nas/zhangyiming
73
+ ln -sfn /actual/database/path /mnt/nas/zhangyiming/database
74
+ ```
75
+
76
+ ## Asset Notes
77
+
78
+ ### Action Expert
79
+
80
+ Expected target path:
81
+
82
+ ```text
83
+ /mnt/nas/zhangyiming/database/ckpt/pretrained/LaST0_Pretrain_AE_chunk16/tfmr
84
+ ```
85
+
86
+ Files:
87
+
88
+ ```text
89
+ config.json
90
+ model.safetensors
91
+ preprocessor_config.json
92
+ processor_config.json
93
+ special_tokens_map.json
94
+ tokenizer.json
95
+ tokenizer_config.json
96
+ ```
97
+
98
+ This directory is used as both the action expert and processor/tokenizer source
99
+ by the MoT2 scripts. Do not replace it with Janus-Pro-1B.
100
+
101
+ ### RLBench
102
+
103
+ Expected target root:
104
+
105
+ ```text
106
+ /mnt/nas/zhangyiming/database/rlbench
107
+ ```
108
+
109
+ The full source `rlbench` directory is included. The key training files used by
110
+ the current scripts include:
111
+
112
+ ```text
113
+ rlbench/train/json/train_action_chunk1_sumpos_lastrot.json
114
+ rlbench/train/json/train_action_chunk1_sumpos_lastrot_statistics.json
115
+ rlbench/train/json/cosmos_text_cache_rlbench_keyframe/
116
+ rlbench/train/images/
117
+ ```
118
+
119
+ The JSON references image paths such as:
120
+
121
+ ```text
122
+ /mnt/nas/zhangyiming/database/rlbench/train/images/...
123
+ ```
124
+
125
+ ### LIBERO
126
+
127
+ Expected target path:
128
+
129
+ ```text
130
+ /mnt/nas/zhangyiming/database/data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual
131
+ ```
132
+
133
+ Included files/directories:
134
+
135
+ ```text
136
+ train_with_atomic_action.json
137
+ train.json
138
+ train_rewritten_prompts.json
139
+ train_statistics.json
140
+ videos/
141
+ cosmos_text_cache/
142
+ cosmos_text_cache_raw_full_concat/
143
+ cosmos_text_cache_rewritten_prompts_raw_full_concat/
144
+ ```
145
+
146
+ The JSON references video paths such as:
147
+
148
+ ```text
149
+ /mnt/nas/zhangyiming/database/data/libero_training_data_last05_lastest/libero_spatial_20hz_224_dual/videos/...
150
+ ```
151
+
152
+ ## Main Training Entrypoints
153
+
154
+ RLBench:
155
+
156
+ ```bash
157
+ cd /mnt/nas/zhangyiming/last05_beta/last05_mot2_action
158
+ bash scripts/train_mot2_rlbench_keyframe.sh
159
+ ```
160
+
161
+ LIBERO:
162
+
163
+ ```bash
164
+ cd /mnt/nas/zhangyiming/last05_beta/last05_mot2_action
165
+ bash scripts/train_mot2.sh
166
+ ```