humancompatible-train documentation

Welcome to the humancompatible-train documentation.

What is humancompatible-train?

humancompatible-train is a PyTorch-based package for constrained optimization, aimed at constrained deep learning tasks. We implement several first-order Lagrangian-based methods for constrained optimization with a PyTorch-based API that allow seamless integration of constraints into the training loop.

Getting Started

• Installation
• Getting Started

Tutorials

• Constrained Optimization Overview
• Basic usage: Fairness
• Tips and Tricks

API reference

• Dual Optimizers
  • ALM
  • iALM
  • PBM
• Utils
  • BalancedBatchSampler