framework-reproducibility

Providing reproducibility in deep learning frameworks


Keywords
framework, tensorflow, gpu, deep-learning, determinism, reproducibility, pytorch, seed, seeder, noise, noise-reduction, variance-reduction, atomics, ngc, gpu-determinism, deterministic-ops, frameworks, gpu-support, d9m, r13y, fwr13y
License
Apache-2.0
Install
pip install framework-reproducibility==0.6.0

Documentation

Framework Reproducibility (fwr13y)

Repository Name Change

The name of this GitHub repository was changed to framework-reproducibility on 2023-02-14. Prior to this, it was named framework-determinism. Before that, it was named tensorflow-determinism.

"In addition to redirecting all web traffic, all git clone, git fetch, or git push operations targetting the previous location[s] will continue to function as if made to the new location. However, to reduce confusion, we strongly recommend updating any existing local clones to point to the new repository URL." -- GitHub documentation

Repository Intention

This repository is intended to:

  • provide documentation, status, patches, and tools related to determinism (bit-accurate, run-to-run reproducibility) in deep learning frameworks, with a focus on determinism when running on GPUs, and
  • provide a tool, and related guidelines, for reducing variance (Seeder) in deep learning frameworks.