gongxinyuu/sngan.pytorch

An unofficial Pytorch implementation of SNGAN, achieving IS of 8.21 and FID of 14.21 on CIFAR-10.

SNGAN.pytorch

An unofficial Pytorch implementation of Spectral Normalization for Generative Adversarial Networks. For official Chainer implementation please refer to https://github.com/pfnet-research/sngan_projection

Our implementation achieves Inception score of 8.21 and FID score of 14.21 on unconditional CIFAR-10 image generation task. In comparison, the original paper claims 8.22 and 21.7 respectively.

Set-up

install libraries:

pip install -r requirements.txt

prepare fid statistic file

mkdir fid_stat

Download the pre-calculated statistics for CIFAR10, fid_stats_cifar10_train.npz, to ./fid_stat.

train

sh exps/sngan_cifar10.sh

test

mkdir pre_trained

Download the pre-trained SNGAN model sngan_cifar10.pth to ./pre_trained. Run the following script:

sh exps/eval.sh

Acknowledgement

1. Inception Score code from OpenAI's Improved GAN (official).
2. FID code and statistics file from https://github.com/bioinf-jku/TTUR (official).
3. The code of Spectral Norm GAN is inspired by https://github.com/pfnet-research/sngan_projection (official).

sngan.pytorch's Issues

I get FID 21 as claimed in paper and not FID 14 as stated in repository.

Our implementation achieves Inception score of 8.21 and FID score of 14.21 on unconditional CIFAR-10 image generation task. In comparison, the original paper claims 8.22 and 21.7 respectively.

I think I can explain the difference. Some people compute FID with training samples while others use validation samples. I used validation samples and got FID the same as the paper (~21), so I'd guess you used training samples which would explain the slightly lower score.

Question 1 Did you compute FID with training or validation samples?

How to set the learning rate adjustment method？

Hello, our research group is using sngan code recently. When training cifar10, what is the parameter LR_ Is decay set to true by default? Thank you for your reply.
parser.add_argument( '--lr_decay', action='store_true', help='learning rate decay or not')

pre-calculated statistics for STL10

Hi,

Do u have any idea how to download the pre-calculated statistics for STL10？like 'stl10_train_unlabeled_fid_stats_48.npz' in 79 line in train.py?