tangger/openpi
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Streamline processor loading logic

Browse Source
This commit is contained in:
Ury Zhilinsky
2024-12-25 15:19:42 -08:00
committed by uzhilinsky
parent fc7b7bc694
commit 06c632b144
4 changed files with 41 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
README.md
View File

@@ -104,11 +104,13 @@ The training config is used to determine which data transformations should be ap
There are also a number of checkpoints that are available as exported JAX graphs, which we trained ourselves using our internal training code. These can be served using the following command:
```bash
uv run scripts/serve_policy.py --env ALOHA policy:exported --policy.dir=s3://openpi-assets/exported/pi0_aloha/model --policy.processor=trossen_biarm_single_base_cam_24dim
uv run scripts/serve_policy.py --env ALOHA policy:exported --policy.dir=s3://openpi-assets/exported/pi0_aloha/model [--policy.processor=trossen_biarm_single_base_cam_24dim]
```
In this case, the data transformations are taken from the default policy and the processor name will be used to determine which norms stats should be used to normalize the transformed data.
For these exported models, norm stats are loaded from processors that are exported along with the model, while data transformations are defined in the corresponding default policy (see `create_default_policy` in [scripts/serve_policy.py](scripts/serve_policy.py)). The processor name is optional, and if not provided, we will do the following:
- Try using the default environment processor name
- Load a processor if there is only one available
- Raise an error if there are multiple processors available and ask to provide a processor name
### Running with Docker: