Focus on your deep learning experiments and forget about (re)writing code. lighter is:

1. Task-agnostic

   Whether you’re working on classification, segmentation, or self-supervised learning, lighter provides generalized training logic that you can use out-of-the-box.

2. Configuration-based

   Easily define, track, and reproduce experiments with lighter’s configuration-driven approach, keeping all your hyperparameters organized.

3. Customizable

   Seamlessly integrate your custom models, datasets, or transformations into lighter’s flexible framework.

 

lighter stands on the shoulder of these two giants:

• MONAI Bundle - Configuration system. Similar to Hydra, but with additional features.
• PyTorch Lightning - Our LighterSystem is based on the PyTorch Lightning LightningModule and implements all the necessary training logic for you. Couple it with the PyTorch Lightning Trainer and you're good to go.

Simply put, lighter = config(trainer + system) 😇

📖 Getting Started

📚 Documentation    🎥 YouTube Channel    👾 Discord Server

Install:

pip install project-lighter

Pre-release (up-to-date with the main branch):

pip install project-lighter --pre

For development:

make setup
make install             # Install lighter via Poetry
make pre-commit-install  # Set up the pre-commit hook for code formatting
poetry shell             # Once installed, activate the poetry shell

💡 Projects

Projects that use lighter:

Project	Description
Foundation Models for Quantitative Imaging Biomarker Discovery in Cancer Imaging	A foundation model for lesions on CT scans that can be applied to down-stream tasks related to tumor radiomics, nodule classification, etc.

Cite

@software{lighter,
author       = {Ibrahim Hadzic and
                Suraj Pai and
                Keno Bressem and
                Hugo Aerts},
title        = {Lighter},
publisher    = {Zenodo},
doi          = {10.5281/zenodo.8007711},
url          = {https://doi.org/10.5281/zenodo.8007711}
}