Skip to content
Branch: master
Find file Copy path
Find file Copy path
2 contributors

Users who have contributed to this file

@rsepassi @cyfra
98 lines (71 sloc) 4.01 KB

How to Contribute

Thanks for thinking about contributing to our library !

Before you start

  • Please accept the Contributor License Agreement (see below)
  • Comment on the issue that you plan to work on so we can assign it to you and there isn't unnecessary duplication of work. If this is your first time contributing, we'll send you an invitation on GitHub to be a contributor; you must accept this invitation here before we can assign you the issue.
  • When you plan to work on something larger (for example, adding new FeatureConnectors), please respond on the issue (or create one if there isn't one) to explain your plan and give others a chance to discuss.
  • If you're fixing some smaller issue - please check the list of pending Pull Requests to avoid unnecessary duplication.

How you can help:

You can help in multiple ways:

  • Adding new datasets and/or requested features (see the issues)
  • Reproducing bugs reported by others: This helps us a lot.
  • Doing code reviews on the Pull Requests from the community.
  • Verifying that Pull Requests from others are working correctly (especially the ones that add new datasets).


Adding a public dataset to tensorflow-datasets is a great way of making it more accessible to the TensorFlow community.

See our Add a dataset doc to learn how to add a dataset.


Methods and classes should have clear and complete docstrings. Most methods (and all publicly-facing API methods) should have an Args: section that documents the name, type, and description of each argument. Argument lines should be formatted as arg_name: (`arg_type`) Description of arg.

References to tfds methods or classes within a docstring should go in backticks and use the publicly accessible path to that symbol. For example `tfds.core.DatasetBuilder`. Doing so ensures that the API documentation will insert a link to the documentation for that symbol.


To ensure that tensorflow-datasets is nice to use and nice to work on long-term, all modules should have clear tests for public members. All tests require:

  • Subclassing tfds.testing.TestCase
  • Calling tf.compat.v1.enable_eager_execution() at the top-level just after the imports. This is to enable testing against TF 2.0.
  • Using the @tfds.testing.run_in_graph_and_eager_modes() decorator for all functionality that touches TF ops. To evaluate Tensor values in a way that is compatible in both Graph and Eager modes, use self.evaluate(tensors) or tfds.as_numpy.
  • End the test file with if __name__ == "__main__": tfds.testing.test_main().

Note that tests for DatasetBuilders are different and are documented in the guide to add a dataset.

Pull Requests

All contributions are done through Pull Requests here on GitHub.

Contributor License Agreement

Contributions to this project must be accompanied by a Contributor License Agreement. You (or your employer) retain the copyright to your contribution, this simply gives us permission to use and redistribute your contributions as part of the project. Head over to to see your current agreements on file or to sign a new one.

You generally only need to submit a CLA once, so if you've already submitted one (even if it was for a different project), you probably don't need to do it again.

Code reviews

All submissions, including submissions by project members, require review. We use GitHub pull requests for this purpose. Consult GitHub Help for more information on using pull requests.

You can’t perform that action at this time.