Contributing

Issues and pull requests are welcome on GitHub.

For a bug report, include the smallest input example that reproduces the issue, the full prediction output, and the Python, PyTorch, NumPy, and Pillow versions involved.

For a code change:

Add or update tests for the behavior you changed.
Run python -m pytest -q.
If you touch preprocessing, weights, or metrics, re-run python examples/verify_metrics.py and include its output — the model card's numbers must always reproduce from the shipped artifacts.
Do not add benchmark claims without the script, data description, and measured output needed to reproduce them.