open-metric-learning

Why we included this project

Most training recipes teach a network to classify and never directly shape how its embeddings sit in space. Metric learning does: it arranges embeddings so similar samples land close together and dissimilar ones drift apart, which is what makes image retrieval, near-duplicate detection, and product search work in practice. OML packages that workflow into a single PyTorch framework. Instead of assembling research code yourself, you get the specialized losses, miners, and samplers that metric learning needs, along with pretrained extractors such as ViT and CLIP. Its pipelines handle feature extraction and post-processing like adaptive thresholding and Siamese reranking, and they report evaluation metrics such as precision-at-k and FNR@FMR. If you already work in PyTorch Lightning, the training modules and DDP support drop in cleanly, and the same code runs on image, text, or audio inputs.