AdversarialNAS-GAN

Code for our CVPR 2020 paper "AdversarialNAS: Adversarial Neural Architecture Search for GANs".

Contributed by Chen Gao, Yunpeng Chen, Si Liu, Zhenxiong Tan, Shuicheng Yan.

Introduction

We propose a novel AdversarialNAS method, which is the first gradient-based fully differentiable NAS method in GAN field and achieves state-of-art performance with much higher efficiency. We also design a large architecture search space for GAN and make it feasible to search in.

Architecture search snapshots

Checklist

Searched architecture of generative model on CIFAR-10 dataset.
Re-trained generative models on CIFAR-10 and STL-10.
Inference codes for evaluation.
Training codes for the searched architectures.
Searching codes.

Getting Started

Installation

Clone this repository.

git clone https://github.com/chengaopro/AdversarialNAS.git

Install pytorch 1.1.0, tensorflow 1.9.0, CUDA 9.0 and corresponding CUDNN via conda.

conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch

conda install tensorflow-gpu==1.9.0 cudnn

Install the requirements via pip.
```
pip install -r requirements.txt
```

Preparing necessary files

Files can be found in Google Drive and Baidu Drive (passcode: bk1m).

Download the pre-calculated statistics to ./fid_stat for calculating the FID.
```
mkdir fid_stat
```
Download the inception model to ./tmp for calculating the IS and FID.
```
mkdir tmp
```

Search

New features

Better performance
Supporting Amended Gradient Estimation

Train and Test the searched architecture.

Training

Train the weights of generative model with the searched architecture (the architecture is saved in ./exps/arch_cifar10/Genotypes/latest_G.npy)
```
sh scripts/train_arch_cifar10.sh
```
or
```
sh scripts/train_arch_stl10.sh
```

Testing

Download the two trained generative models (Google Drive and Baidu Drive (passcode: bk1m)) to ./exps/arch_train_cifar10/Model and ./exps/arch_train_stl10/Model respectively.
```
mkdir -p exps/arch_train_cifar10/Model exps/arch_train_stl10/Model 
```

Test the trained model

sh scripts/test_arch_cifar10.sh

sh scripts/test_arch_stl10.sh

Citation

Please consider citing our paper in your publications if it helps your research. The following is a BibTeX reference. The BibTeX entry requires the url LaTeX package.

@InProceedings{Gao_2020_CVPR,
author = {Gao, Chen and Chen, Yunpeng and Liu, Si and Tan, Zhenxiong and Yan, Shuicheng},
title = {AdversarialNAS: Adversarial Neural Architecture Search for GANs},
booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}

License

AdversarialNAS is released under the MIT license. See LICENSE for additional details.

Acknowledge

Some of the codes are built upon improved-gan, AutoGAN and DARTS. Thanks them for their great works!