Research paper code contribution

How to contribute a research paper code implementation

We encourage researchers to publish new state-of-the-art machine learning models to the TensorFlow Model Garden.

We provide a 📝 README template that describes the information required for publishing a new code implementation.

To contribute a new research paper code, please provide your plans using GitHub issues in this repository before making any pull requests.

Requirements

We want to ensure research code implementations from contributors are high-quality and well-documented.

Your contributions must meet the following requirements to be accepted to the TensorFlow Model Garden repository.

Directory	Requirements
official	• Models implemented using the modelling libraries provided by the Model Garden • Pass the TensorFlow code usability review process • Provide baseline results • Support distributed traininig on GPUs and TPUs • Reasonable performance on GPUs and TPUs • Support provided from TensorFlow
research	• Models implemented in TensorFlow 2 by paper authors at Google or external contributors • Provide baseline results • Reasonable performance on GPUs or TPUs • Support provided from maintainers or TensorFlow
community	• Models implemented in TensorFlow 2 by external contributors • Reproduce the paper results

Model selection

• A model from the paper accepted at top machine learning venues
• A state-of-the-art model from a pre-publication available at arXiv

Model accuracy and performance

• Should be able to reproduce the same results in a published paper
• Should provide reasonable out-of-box performance
  • Should have accuracy and performance test results on GPUs or TPUs

Pre-trained models

• Pre-trained models in TensorFlow SavedModel format should be published to TensorFlow Hub.

Documentation

• Should use Read the Docs for hosting documentation
  • Documentation can be automatically generated from your repository.
  • Read the Docs Template