pytorch-metric-learning

Why we included this project

Learning to tell similar items apart is the task behind face verification, product recognition, person re-identification, and image retrieval, and this library supplies the standard machinery for it without tying you to a new framework. Its triplet and pair-based losses, hard-pair miners, distance functions, and reducers are modular components that drop into an existing PyTorch training loop. You can start with a plain TripletMarginLoss and swap in cosine distances, threshold reducers, or regularizers one at a time as your needs change. Self-supervised setups and a cross-batch memory queue for MoCo-style training are covered too, and the repo ships trainers, testers, and an accuracy calculator for a complete workflow. Teams already comfortable in PyTorch get a practical way to build embedding models without reimplementing triplet mining themselves.