act-baseline-seed42

備份自 ID Lab lerobot-vla-fbagent 專案，含完整中間 checkpoint。

訓練資訊

項目	值
Policy Type	act
Dataset	HuggingFaceVLA/libero
Max Steps	10,000
Checkpoints	2 個
Seed	?
Batch Size	?
Learning Rate	?
Num Workers	?
Chunk Size	100
Git Commit	c92e49c
Source Dir	18-19-28_act

復現指令

git clone --recurse-submodules git@github.com:KunHsiang/lerobot-vla-fbagent.git
cd lerobot-vla-fbagent
git checkout c92e49c
uv sync

HF_HUB_ENABLE_HF_TRANSFER=1 .venv/bin/python scripts/train_with_fb.py \
    --policy.type=act \
    --policy.repo_id=kunhsiang/act-baseline-seed42 \
    --dataset.repo_id=HuggingFaceVLA/libero \
    --training.seed=? \
    --training.batch_size=? \
    --training.learning_rate=? \
    --training.num_workers=?

Checkpoint 結構

checkpoints/
  005000/          # step 5,000
  010000/          # step 10,000
  ...
  010000/          # final (10,000 steps)

每個 step 目錄包含:

• pretrained_model/model.safetensors — 模型權重
• pretrained_model/config.json — 模型設定
• pretrained_model/train_config.json — 完整訓練設定
• training_state/optimizer_state.safetensors — Optimizer 狀態（可 resume）
• training_state/rng_state.safetensors — RNG 狀態（可 resume）

Resume 訓練

.venv/bin/python scripts/train_with_fb.py \
    --policy.type=act \
    --resume=true \
    --policy.pretrained_path=checkpoints/010000/pretrained_model