Skip to content

Latest commit

 

History

History
71 lines (60 loc) · 4.26 KB

CONTRIBUTING_TECHNIQUE.md

File metadata and controls

71 lines (60 loc) · 4.26 KB
Raw

Guidelines for Contributing a New Technique

The following are guidelines for contributing a complete technique to the toolkit, for example, connection pruning.

We will be evolving these guidelines to make the process more effective, and as we receive more contributions. For example, we are working on creating a repository of training scripts for different models and tasks aimed at simplifying technique validation (issue), making the project contributor-friendly (issue), and having reproducible results.

Contribution Process

  1. Before anything else, provide a proposal RFC that focuses on what would motivate a user to use this technique, rather than on the specifics of the API. If approved, you will be matched with a sponsor, following the general TensorFlow RFC process. Consider the following:

    • The end to end user story. It's encouraged to consider how the technique fits with the other parts of the toolkit.
    • Experiment results for real world models that support the user story, with not only accuracy but also deployment metrics (e.g. model storage space, latency, memory, to mention a few).
    • Contributing requires ownership of the technique. Select the type of ownership. TODO(tfmot): link to ownership RFC.
    • TODO(tfmot): link to sample pruning proposal RFC.

  2. In parallel:

    • Provide a design RFC under model-optimization/community/rfcs. See the TensorFlow process for details.
      • When reasonable, the API should strive for similarity with existing techniques to provide the best user experience.
      • TODO(tfmot): link to sample pruning design RFC.
    • Start prototyping and building the library in a fork of TFMOT.

  3. Upon approval of design RFC, upstream existing code into TFMOT and continue development there.

  4. Implement the design RFC. In parallel, create a training script to reproduce the results in the RFC. This script will serve as an example to users and come with tests to ensure that results continue to be reproducible. See this page for details.

  5. Documentation and tutorials:

  6. Packaging and release:

    • Releases are managed by the TensorFlow Model Optimization team. Work with them to produce and test pip package as well as generate API docs. To generate API docs locally, see instructions here.
    • Days before the release date:
      • The API docs will not be checked in, but should be tested and ready to submit.
      • The other documentation will be be checked in, but not linked in the navigation bar. The colabs should use a stable TFMOT release, as opposed to a test release.

  7. Collaborative blog post (optional)