KevinMusgrave/pytorch-metric-learning

Provides 9 independent modules—including loss functions (TripletMarginLoss, ArcFaceLoss, etc.), hard pair miners, distance metrics, and reducers—that compose into customizable training pipelines. Automatically converts between triplet and pair representations, supports self-supervised learning via augmentation pairs, and includes built-in trainers and evaluation utilities for metric learning workflows. Integrates seamlessly with standard PyTorch training loops and includes pre-built dataset loaders for common benchmarks like CUB-200 and Stanford Online Products.

6,312 stars and 5,671,502 monthly downloads. Used by 5 other packages. No commits in the last 6 months. Available on PyPI.