Verified project record

KevinMusgrave/pytorch-metric-learning

pytorch-metric-learning is a modular library that simplifies deep metric learning by providing losses, miners, trainers, and testers as independent components. It addresses the complexity of building and combining these elements from scratch so that they can be used in a complete train and test workflow.

Project overview

The library reduces the barrier to entry for deep metric learning through nine independently combinable modules, making it straightforward to customize loss functions with distances, reducers, and regularizers.

Project type: Model Development · Data Processing · Infrastructure
Deployment: Refer to project documentation
License: MIT

Best for

Developers who need a modular and extensible library for deep metric learning with independently combinable components.
Researchers and AI engineers working on self-supervised learning or custom loss function design.

Key capabilities

Nine modules including losses, miners, trainers, and testers that can be used independently or combined for a complete train and test workflow.
Various loss functions such as TripletMarginLoss that can be customized using distances, reducers, and regularizers.
A SelfSupervisedLoss wrapper for self-supervised learning, including support for MoCo-style momentum encoder queues via CrossBatchMemory.
A Datasets module for easy downloading of common datasets such as CUB200, Cars196, and INaturalist 2018.
An AccuracyCalculator to directly compute the accuracy of an embedding space.

Limitations and risks

For library versions >= v0.9.90, PyTorch version >= 1.6 is required.

Getting started

Install the library using pip install pytorch-metric-learning. The installation is straightforward and the package is also available via conda.

Evidence and sources

GitHub project description: The easiest way to use deep metric learning in your application. Modular, flexible, and extensible. Written in PyTorch.
Release: v2.9.0
README: Added the [Datasets](https://kevinmusgrave.github.io/pytorch-metric-learning/datasets) module for easy downloading of common datasets:
README: This library contains 9 modules, each of which can be used independently within your existing codebase, or combined together for a complete train/test workflow.
README: To compute the loss in your training loop, pass in the embeddings computed by your model, and the corresponding labels.

Reviewed at: 2026-07-19T18:57:47Z