Skip to content

Research paper code contribution

Jaeyoun Kim edited this page Aug 3, 2020 · 12 revisions

How to contribute a research paper code implementation

help wanted:paper implementation

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

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 • Provide a model implemented in TensorFlow 2
• Use the modelling libraries provided by the Model Garden
• Provide baseline results
• Support distributed training on GPUs and TPUs
• Reasonable performance on GPUs and TPUs
• Need a SLA (Service Level Agreement) for community support
• Pass the TensorFlow code usability review process
research • Provide a model implemented in TensorFlow 2
• Use the modelling libraries provided by the Model Garden for supported ML tasks
• Provide baseline results
• Reasonable performance on GPUs or TPUs
• Need a SLA (Service Level Agreement) for community support
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 or
  • 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

  • Use the README template that describes the information required for publishing a new code implementation.
  • We also recommend to use Read the Docs for hosting documentation.

Note: Exceptions can be made case by case basis.