add end-to-end pi0 reproduction notebook
This commit is contained in:
@@ -44,6 +44,22 @@ python -m pip install -e '.[dev]'
|
||||
pytest -q
|
||||
```
|
||||
|
||||
## 一键 Notebook 复现
|
||||
|
||||
完整的论文链接、架构拆解、数据准备、过拟合训练、checkpoint 重载、推理指标和轨迹图都在:
|
||||
|
||||
[`notebooks/pi0_minimal_reproduction.ipynb`](notebooks/pi0_minimal_reproduction.ipynb)
|
||||
|
||||
```bash
|
||||
python3.12 -m venv .venv
|
||||
source .venv/bin/activate
|
||||
python -m pip install -U pip
|
||||
python -m pip install -e '.[notebook,dev]'
|
||||
python -m jupyter lab notebooks/pi0_minimal_reproduction.ipynb
|
||||
```
|
||||
|
||||
Notebook 默认可直接 Run All:优先读取 `PI0_SOURCE_MANIFEST` 指定的已处理数据,其次尝试相邻 Alpamayo 项目,最后自动生成一个自包含 synthetic 样本。可通过 `PI0_PROFILE=smoke` 将训练缩短到 5 步,仅检查执行链路;默认 `full` 使用 600 步完成过拟合验证。
|
||||
|
||||
`tests/test_flow.py` 验证直线概率路径 `x_t=(1-t)·noise+t·action`、目标速度 `action-noise`、MSE 损失、随机可复现性和 Euler 积分。
|
||||
|
||||
## 过拟合训练与效果图
|
||||
|
||||
Reference in New Issue
Block a user