This repository contains the code for CVPR 2020 paper Neural Architecture Search for Lightweight Non-Local Networks. This paper presents a lightweight non-local block and automatically searched state-of-the-art non-local networks for mobile vision. We also provide pytorch implementation here.

If you use the code, please cite:

@inproceedings{li2020neural,
title={Neural Architecture Search for Lightweight Non-local Networks},
author={Li, Yingwei and Jin, Xiaojie and Mei, Jieru and Lian, Xiaochen and Yang, Linjie and Xie, Cihang and Yu, Qihang and Zhou, Yuyin and Bai, Song and Yuille, Alan},
booktitle={CVPR},
year={2020}
}

TensorFlow 1.14.0

tensorpack 0.9.8 (for dataset loading)

Download the AutoNL-L-77.7.zip and AutoNL-S-76.5.zip pretrained models. Unzip and place them at the root directory of the source code.

Download and place the ImageNet validation set at $PATH_TO_IMAGENET/val.

python eval.py --model_dir=AutoNL-S-76.5 --valdir=$PATH_TO_IMAGENET/val --arch=AutoNL-S-76.5/arch.txt
python eval.py --model_dir=AutoNL-L-77.7 --valdir=$PATH_TO_IMAGENET/val --arch=AutoNL-L-77.7/arch.txt

The last printed line should read:

Test: [50000/50000]     Prec@1 77.7     Prec@5 93.7

for AutoNL-L, and

Test: [50000/50000]     Prec@1 76.5     Prec@5 93.1

for AutoNL-S.

Part of code comes from single-path-nas, mnasnet and ImageNet-Adversarial-Training.

If you encounter any problems or have any inquiries, please contact us at yingwei.li@jhu.edu.