(ECCV2022) EAGAN: Efficient Two-stage Evolutionary Architecture Search for GANs(PDF)(slides)

Guohao Ying$\dagger$, Xin He$\dagger$, Bin Gao, Bo Han, and Xiaowen Chu $\ddagger$

$$\dagger: \text{Equal contribution}. $$ $$\ddagger: \text{Corresponding author (xwchu@ust.hk).}$$

eccv2022.mp4

Introduction

This is the official implementation of EAGAN: Efficient Two-stage Evolutionary Architecture Search for GANs. We introduce a novel NAS framework, namely EAGAN, to alleviate the instability when searching GANs. Our EAGAN decouples the search into two stages, where stage-1 searches G with a fixed D and adopts the many-to-one training strategy, and stage-2 searches D with the optimal G found in stage-1 and adopts the one-to-one training strategy and the weight-resetting strategy to enhance the stability of GAN training.

The framework of the proposed method:

Installation

Download

git clone https://github.com/marsggbo/EAGAN
cd EAGAN

Environment

We recommend using Anaconda to manage the python environment:

conda create -n EAGAN python=3.8
conda activate EAGAN
pip install -r requirements.txt

Datasets

In our work, we use CIFAR-10 and STL-10 datasets. The default path for both datasets is ~/datasets/cifar10 and ~/datasets/stl10. Therefore, you should move or download both datasets in this path.

Searching

1. Search generator

bash ./scripts/train_search_gen.sh

or run the following commands

python search_gen_arch.py \
--gpu_ids 0 \
--num_workers 16 \
--gen_bs 80 \
--dis_bs 40 \
--dataset cifar10 \
--bottom_width 4 \
--img_size 32 \
--max_epoch_G 120 \
--n_critic 1 \
--arch arch_cifar10 \
--draw_arch False \
--latent_dim 120 \
--gf_dim 256 \
--df_dim 128 \
--g_lr 0.0002 \
--d_lr 0.00004 \
--beta1 0.0 \
--beta2 0.9 \
--init_type xavier_uniform \
--val_freq 5 \
--num_eval_imgs 5000 \
--exp_name arch_train_cifar10 \
--data_path ~/datasets/cifar10

• genotypes_exp: specify the genotype path of initial discriminator.
  • exps/cifar10_D.npy is used by default.
  • if you want to specify a different initial discriminator, you can modify dis_genotype variable in search_gen_arch.py, e.g., dis_genotype=np.load('./exps/cifar10_D.npy')

2. Search discriminator

bash ./scripts/train_search_dis.sh

3. Searched architecture

• ./exps/best_G.npy is the best searched generator genotype that can be trained on cifar10 and stl10 datasets, presented in the paper Fig.3.
• ./exps/cifar10_D.npy and ./exps/stl10_D.npy are the best searched discriminator genotypes for cifar10 and stl10 datasets respectively, and their details can be found in the paper Fig.4.

Fully train

1. Fully train GAN on cifar10

bash ./scripts/train_arch_cifar10.sh

or run the following commands

python fully_train_arch.py \
--gpu_ids 0 \
--num_workers 16 \
--gen_bs 80 \
--dis_bs 40 \
--dataset cifar10 \
--bottom_width 4 \
--img_size 32 \
--max_epoch_G 120 \
--n_critic 5 \
--arch arch_cifar10 \
--draw_arch False \
--genotypes_exp cifar10_D1.npy \
--latent_dim 120 \
--gf_dim 256 \
--df_dim 128 \
--g_lr 0.0002 \
--d_lr 0.0002 \
--beta1 0.0 \
--beta2 0.9 \
--init_type xavier_uniform \
--val_freq 5 \
--num_eval_imgs 50000 \
--exp_name arch_train_cifar10 \
--data_path ~/datasets/cifar10

2. Fully train GAN on stl10

bash ./scripts/train_arch_stl10.sh

3. Test

bash ./scripts/test_arch_cifar10.sh
bash ./scripts/test_arch_stl10.sh

Citation

If you find this work useful for your research, please kindly cite our paper:

@inproceedings{ECCV2022EAGAN,
  title={EAGAN: Efficient Two-stage Evolutionary Architecture Search for GANs},
  author={Guohao Ying, Xin He, Bin Gao, Bo Han, and Xiaowen Chu.},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2022}
}

Acknowledgements

Thanks NVIDIA AI TECHNOLOGY CENTER (NVAITC) for providing GPU clusters to support our work.