Lightly
A Python computer vision framework for self-supervised learning, offering modular loss functions and model heads with PyTorch and PyTorch Lightning-based examples for many published methods.
Lightly, packaged as LightlySSL, is a computer vision framework for self-supervised learning. It exposes low-level building blocks such as loss functions and model heads in a PyTorch-like style, supports custom backbone models for pretraining, and supports distributed training through PyTorch Lightning.
It includes sample code and reference implementations for many published self-supervised methods, including MoCo, SimCLR, BYOL, SwaV, DINO, VICReg, MAE, iBOT and SimMIM, with PyTorch, PyTorch Lightning and distributed PyTorch Lightning variants for each. A commercial version adds Docker support and single-command pretraining for embedding, classification, detection and segmentation, and a separate hosted product, LightlyStudio, provides data curation, labeling and active learning on top of the open-source framework.
It is licensed under MIT.