A modern PyTorch Lightning reimplementation of Differentiable Architecture Search
This project is a fork of the original DARTS repository. The goal is to enable further research and development in a field of Neural Architecture Search, through a gradual upgrade and improvement of the original Differentiable Architecture Search implementation
The modifications include:
- Upgrading code to PyTorch 2.x
- Rewriting the code to use PyTorch Lightning for training and architecture search
- Make things optimized for modern hardware
- Fix bugs
The scope is now limited to Convolutional Neural Networks (CNN) and once the reimplementation is done, I will expand it to other kinds of neural nets
The original paper:
DARTS: Differentiable Architecture Search
Hanxiao Liu, Karen Simonyan, Yiming Yang.
arXiv:1806.09055.
The algorithm is based on continuous relaxation and gradient descent in the architecture space. It is able to efficiently design high-performance convolutional architectures for image classification (on CIFAR-10 and ImageNet) and recurrent architectures for language modeling (on Penn Treebank and WikiText-2). Only a single GPU is required.
- Python >= 3.10
- PyTorch >= 2.4.1
- PyTorch Lightning >= 2.4.0
If you use this project in your research, please use the following citation:
@misc{Maczan_DARTS_2024,
title = "Lightning Differentiable Architecture Search: A modern PyTorch Lightning reimplementation of Differentiable Architecture Search (DARTS)",
author = "Maczan, Jędrzej Paweł",
howpublished = "\url{https://github.com/jmaczan/darts}",
year = 2024,
publisher = {GitHub}
}Please strongly consider to cite the original DARTS paper as well:
@article{liu2018darts,
title={DARTS: Differentiable Architecture Search},
author={Liu, Hanxiao and Simonyan, Karen and Yang, Yiming},
journal={arXiv preprint arXiv:1806.09055},
year={2018}
}Apache License 2.0
Original DARTS paper and implementation: Hanxiao Liu, Karen Simonyan, Yiming Yang
Lightning Differentiable Architecture Search: Jędrzej Maczan