Repolode
Submit
lightly-ai

lightly-ai/lightly

A Python library for self-supervised learning on images, offering PyTorch building blocks for training vision models without labels.

View on GitHubWebsite
Stars
3,829
Forks
379
Open issues
119
Last commit
8 days ago

What is it?

Lightly SSL is a computer vision framework for self-supervised learning, a technique that trains image models on unlabeled data. It targets developers who want to pretrain a backbone on their own image collection instead of relying only on ImageNet-style pretrained weights.

The library is written in a PyTorch-like style and exposes low-level building blocks such as loss functions, projection and prediction heads, and data transforms. It includes implementations and examples for many published methods, including MoCo, SimCLR, BYOL, DINO, DINOv2, MAE and VICReg, with PyTorch and PyTorch Lightning examples. Custom backbone models are supported, and distributed multi-GPU training is available through PyTorch Lightning.

Installation is via pip, and the project is MIT-licensed. The maintainers also offer separate commercial products for data curation and labeling, which are not part of this package.

Key features

  • 01Modular building blocks such as loss functions and model heads
  • 02Implementations of many self-supervised methods including SimCLR, BYOL, DINO and MAE
  • 03Support for custom backbone models
  • 04PyTorch and PyTorch Lightning examples
  • 05Distributed training with gathered losses and synchronized BatchNorm

Use cases

  • Pretraining a vision backbone on a domain-specific unlabeled image dataset
  • Comparing different self-supervised methods on the same data
  • Building custom SSL research experiments with swappable losses and heads
  • Training on multiple GPUs with PyTorch Lightning

Things to know

  • Docker support and single-command pretrained models are only in the commercial version.

In collections

Computer vision libraries & tools

Topics

#computer-vision#self-supervised-learning#pytorch#deep-learning#contrastive-learning#embeddings#contributions-welcome#hacktoberfest#machine-learning

FAQ

What is lightly?
A Python library for self-supervised learning on images, offering PyTorch building blocks for training vision models without labels. Lightly SSL is a computer vision framework for self-supervised learning, a technique that trains image models on unlabeled data. It targets developers who want to pretrain a backbone on their own image collection instead of relying only on ImageNet-style pretrained weights.
Is lightly free to use? What license does it have?
lightly is open source under the MIT license. Check the license terms before using it in commercial products.
Is lightly still maintained?
Actively maintained. The last commit to lightly was on Oct 30, 2026.
What can I use lightly for?
Pretraining a vision backbone on a domain-specific unlabeled image dataset Comparing different self-supervised methods on the same data Building custom SSL research experiments with swappable losses and heads Training on multiple GPUs with PyTorch Lightning
Where can I get lightly?
The source code and installation instructions are on GitHub: https://github.com/lightly-ai/lightly

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects