lightly: a modular computer vision framework for self-supervised pre-training

What it solves

LightlySSL is a computer vision framework designed to make self-supervised learning (SSL) more accessible and easier to implement. It allows developers to train models on unlabeled data, reducing the dependency on massive, manually labeled datasets for computer vision tasks.

How it works

Built on PyTorch and PyTorch Lightning, the framework provides a modular set of building blocks—including loss functions, model heads, and data transforms—that can be combined to implement various SSL algorithms. It supports a wide array of pre-training methods and allows users to integrate custom backbone models for feature extraction.

Who it’s for

It is intended for machine learning engineers and researchers working with computer vision who want to implement self-supervised pre-training for tasks like classification, detection, and segmentation.

Highlights

• Extensive Model Support: Includes implementations of popular SSL models such as SimCLR, MoCo, DINO, BYOL, MAE, and LeJEPA.
• Modular Design: Exposes low-level components like loss functions and projection heads for flexible model construction.
• Symmetry with PyTorch: Written in a PyTorch-like style for a seamless developer experience.
• Distributed Training: Native support for distributed training via PyTorch Lightning.