---

name: contribution-and-testing description: "Use for TorchGeo repository edits, style/type conventions, adding datasets/models/tasks/transforms, fake fixtures, docs updates, and targeted verification commands." metadata: disco-role: operating disable-model-invocation: true license: MIT

TorchGeo contribution and testing

Use this sub-skill when modifying the TorchGeo repository rather than only using the package.

Required style conventions

• New Python files must start with:

# Copyright (c) TorchGeo Contributors. All rights reserved.
# Licensed under the MIT License.

• Ruff style uses single quotes and isort ordering: stdlib, third-party, then local imports.
• All functions need type annotations. Prefer built-in collection types and X | Y unions.
• Use Google-style docstrings with Args, Returns, Raises, and Warns sections as needed.
• Avoid broad Exception catches. Use TorchGeo dataset errors such as DatasetNotFoundError, RGBBandsMissingError, or DependencyNotFoundError where applicable.

New dataset checklist

1. Add torchgeo/datasets/<dataset>.py with the license header.
2. Import it in torchgeo/datasets/__init__.py.
3. Add generated/minimal fake data under tests/data/<dataset>/.
4. Add tests/datasets/test_<dataset>.py with enough coverage for initialization, __getitem__, plotting, missing data, checksum/download behavior, and split handling as applicable.
5. Update docs/api/datasets.rst and the relevant dataset CSV under docs/api/datasets/.
6. Add or update a datamodule and datamodule test if the dataset has a standard training workflow.

Targeted commands

From the repository root:

ruff format && ruff check && ty check && prettier --write .
pytest --cov=torchgeo tests/
pytest --cov=torchgeo.datasets tests/datasets/test_<dataset>.py
pytest --cov=torchgeo.datasets tests/datasets/test_<dataset>.py::TestDataset::test_getitem
pytest -m "" --cov=torchgeo tests/  # includes slow tests, use only when approved

For docs, install docs dependencies and run:

cd docs && make clean && make html

Testing practices

• Use fake data only; never commit real dataset samples.
• Mark downloads/network as @pytest.mark.slow and respect the default socket-disabled test settings.
• Use pytest.importorskip for optional dependencies.
• Close matplotlib figures in plot tests.
• Keep files reasonably small and separate logical changes.

Read next

• reference for review checklists and native test maps.