The implementation of AutoGAN: Neural Architecture Search for Generative Adversarial Networks.
We've desinged a novel neural architecture search framework for generative adversarial networks (GANs), dubbed AutoGAN. Experiments validate the effectiveness of AutoGAN on the task of unconditional image generation. Specifically, our discovered architectures achieve highly competitive performance on unconditional image generation task of CIFAR-10, which obtains a record FID score of 12.42, a competitive Inception score of 8.55.
RNN controller:
Search space:
Discovered network architecture:
Unconditional image generation on CIFAR-10.
Unconditional image generation on STL-10.
python >= 3.6
pip install -r requirements.txt
mkdir fid_stat
Download the pre-calculated statistics
(Google Drive) to ./fid_stat
.
sh exps/autogan_cifar10_a.sh
Run the following script:
python test.py \
--dataset cifar10 \
--img_size 32 \
--bottom_width 4 \
--model autogan_cifar10_a \
--latent_dim 128 \
--gf_dim 256 \
--g_spectral_norm False \
--load_path /path/to/*.pth \
--exp_name test_autogan_cifar10_a
Pre-trained models are provided (Google Drive).
If you find this work is useful to your research, please cite our paper:
@InProceedings{Gong_2019_ICCV,
author = {Gong, Xinyu and Chang, Shiyu and Jiang, Yifan and Wang, Zhangyang},
title = {AutoGAN: Neural Architecture Search for Generative Adversarial Networks},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {Oct},
year = {2019}
}
- Inception Score code from OpenAI's Improved GAN (official).
- FID code and CIFAR-10 statistics file from https://github.com/bioinf-jku/TTUR (official).