Framework Reproducibility: Seeder

Introduction

Seeder (implemented in the seeder sub-module) generates deterministic seeds to reduce variance across training (and inference) runs. This reduces the number of runs needed to catch regressions without changing the underlying algorithms used (from nondeterministic to deterministic). It supports suspend/resume functionality.

Seeder is an experimental feature.

Installation

pip install framework-reproducibility --upgrade

Frameworks Supported

See the instructions specific to the framework you're using:

• PyTorch
• TensorFlow2