Awesome

Shapley-NAS

This is the official pytorch implementation for the paper: Shapley-NAS: Discovering Operation Contribution for Neural Architecture Search, which is accepted by CVPR2022. This repo contains the implementation of architecture search and evaluation on CIFAR-10 and ImageNet using our proposed Shapley-NAS.

intro

Quick Start

Prerequisites

python>=3.5
pytorch>=1.1.0
torchvision>=0.3.0

Usage

Architecture Search on CIFAR-10

To search CNN cells on CIFAR-10, please run

export CUDA_VISIBLE_DEVICES=0
python -W ignore train_search.py    \
--batch_size 256    \
--shapley_momentum 0.8  \
--save cifar10_shapley    \
--data /path/to/cifar10

or simply use the command

bash run/search_cifar10_shapley.sh

Architecture Search on ImageNet

To search CNN cells on ImageNet, please run

export CUDA_VISIBLE_DEVICES=0,1,2,3
python -W ignore train_search_imagenet.py    \
--batch_size 1024    \
--save imagenet_shapley    \
--shapley_momentum 0.8  \
--data /path/to/imagennet  \

or simply use the command

bash run/search_imagenet_shapley.sh

Architecture Evaluation on CIFAR-10

To evaluate the derived architecture on CIFAR-10, please run

export CUDA_VISIBLE_DEVICES=0
python -W ignore train.py   \
--data /path/to/cifar10 \
--save train_cifar10   \
--auxiliary \
--cutout    \

or simply use the command

bash run/train_cifar10.sh

Architecture Evaluation on ImageNet

To evaluate the derived architecture on ImageNet, please run

export CUDA_VISIBLE_DEVICES=0,1,2,3
python -W ignore train_imagenet.py \
 --tmp_data_dir /path/to/imagenet \
 --save train_imagenet \
 --workers 16   \
 --auxiliary \
 --note imagenet_shapley    \

or simply use the command

bash run/train_imagenet.sh

Citation

Please cite our paper if you find it useful in your research:

@InProceedings{Xiao_2022_CVPR,
    author    = {Xiao, Han and Wang, Ziwei and Zhu, Zheng and Zhou, Jie and Lu, Jiwen},
    title     = {Shapley-NAS: Discovering Operation Contribution for Neural Architecture Search},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2022}
}

Acknowledgements

We thank the authors of following works for opening source their excellent codes.