This is the official pytorch implementation for the paper: EG-NAS: Neural Architecture Search with Fast Evolutionary Exploration, which is accepted by AAAI2024. This repo contains the implementation of architecture search and evaluation on CIFAR-10 and ImageNet using our proposed EG-NAS.
- python>=3.5
- pytorch>=1.1.0
- torchvision>=0.3.0
- pip install cmaes
To search CNN cells on CIFAR-10, please run
export CUDA_VISIBLE_DEVICES=0
python train_search.py \
--batch_size 256 \
--data /path/to/cifar10
To search CNN cells on ImageNet, please run
export CUDA_VISIBLE_DEVICES=0,1,2,3
python train_search_imagenet.py \
--batch_size 1024 \
--data /path/to/imagennet \
To evaluate the derived architecture on CIFAR-10, please run
export CUDA_VISIBLE_DEVICES=0
python train.py \
--data /path/to/cifar10 \
--save train_cifar10 \
--auxiliary \
--cutout \
To evaluate the derived architecture on ImageNet, please run
export CUDA_VISIBLE_DEVICES=0,1,2,3
python train_imagenet.py \
--tmp_data_dir /path/to/imagenet \
--save train_imagenet \
--workers 16 \
--auxiliary \
--note imagenet_shapley \
Please cite our paper if you find it useful in your research:
@article{Cai_Chen_Liu_Ling_Lai_2024,
title={EG-NAS: Neural Architecture Search with Fast Evolutionary Exploration},
volume={38},
url={https://ojs.aaai.org/index.php/AAAI/article/view/28993},
DOI={10.1609/aaai.v38i10.28993}, number={10},
journal={Proceedings of the AAAI Conference on Artificial Intelligence},
author={Cai, Zicheng and Chen, Lei and Liu, Peng and Ling, Tongtao and Lai, Yutao},
year={2024},
month={Mar.},
pages={11159-11167}
}
We thank the authors of following works for opening source their excellent codes.