lightly
A python library for self-supervised learning on images.
What is lightly?
A python library for self-supervised learning on images.
Overview of lightly
What it is
LightlySSL is a Python library for self-supervised learning on images, developed by lightly-ai. It is a computer vision framework that also supports training latent world models. The library is open-source under the MIT license. A commercial version with additional features such as Docker support and pretrained models for embedding, classification, detection, and segmentation is available by contacting sales. The LightlyStudio platform, built on top of the library, provides tools for active learning and data curation and is free to get started.
Key features
- Modular framework exposing low-level building blocks such as loss functions and model heads.
- Easy to use with a PyTorch-like style.
- Supports custom backbone models for self-supervised pre-training.
- Support for distributed training using PyTorch Lightning.
- Supports a wide range of self-supervised learning models including MoCo, SimCLR, BYOL, SwaV, DINO, MAE, DINOv2, and more, as well as world models like LeWM.
- Includes world model training (LeWM) that learns from raw frames and actions and plans by imagining outcomes.
- Provides tutorials and example scripts for various use cases.
Use cases
- Self-supervised learning on images for representation learning.
- Training world models to understand how a scene responds to actions.
- Data curation and active learning through the LightlyStudio platform.
- On-device deep learning, for example on ARM microcontrollers.
- Pretraining backbones for downstream tasks such as classification, detection, and segmentation (commercial version).
Getting started
Lightly requires Python 3.8+ and is recommended to be installed in a Linux or OSX environment. Dependencies include PyTorch>=1.11.0, Torchvision>=0.12.0, and PyTorch Lightning>=1.7.1. Install the library from PyPI: pip3 install lightly For world model support (LeWM), install with: pip install 'lightly[timm]' See the repository README for a full code example using SimCLR or other models.
License and pricing
The library is open-source under the MIT license. A commercial version with additional features is available by contacting [email protected]. The LightlyStudio platform is free to get started.
Based on the project's public README and repository metadata.