ComGAN PyTorch Implementation.

This is a implementation of our work ComGAN[1] built upon StudioGAN framework[2]. Our paper is avaiable at here.

To Get Started

Please read the StudioGAN documentation at here and finish the basic set up.
After all, you can run, for example, SComGAN by executing the following command.

!python src/main.py -t -hdf5 -l -metrics is fid prdc -ref test --num_workers 2 --save_freq 2000 --print_freq 100 \
    -cfg src/configs/JointGAN/CIFAR10/joint-dcgan.yaml \
    -data ./cifar10 -save . --post_resizer friendly --eval_backbone InceptionV3_tf

For Tiny-ImageNet, please run the following command in the directory of PyTorch-StudioGAN in order to download the dataset.

!wget http://cs231n.stanford.edu/tiny-imagenet-200.zip
!wget https://gist.github.com/rl-max/58825fd2df2ca82fb4693032b3077ea5/raw -O process_tiny_imagenet.sh
!bash process_tiny_imagenet.sh
!mv tiny-imagenet-200/val tiny-imagenet-200/valid
!rm tiny-imagenet-200.zip

Then, excute this.

!python src/main.py -t -hdf5 -l -metrics is fid prdc -ref valid --num_workers 2 --save_freq 2000 --print_freq 100 \
    -cfg src/configs/JointGAN/Tiny_ImageNet/joint-wgan-gp.yaml \
    -data ./tiny-imagenet-200 -save . --post_resizer friendly --eval_backbone InceptionV3_tf

Configurations

To run other losses, corresponding files can be found under src/configs/JointGAN/CIFAR10/(e.g. Use "joint-lsgan.yaml" for least square loss) and similarly for Tiny-ImageNet. For the arguments of the above commmand such as -metric, -t, and -eval_backbone, please refer to the StudioGAN documentation.

In each of yaml files, you will have the following options that are components of ComGAN.

• MODEL.jointgan_arch: This determines which architecture to use for ComGAN. Options are provided among [rgan, ragan, concat].
• LOSS.jointgan_object: This determines whether to use real data, fake data, or same sample for comparative sample[1], each of which is denoted by [r, f, s].
• LOSS.apply_reg: This determines the equality regularization[1] to be applied. The available options are [N/A, l1, l1_mean, l2, l2_mean, logistic, logistic_mean, hf]. We recommend using "l2_mean" for RGAN/RaGAN and "l2" for concat architecture.

References

[1] Haeone Lee. ComGAN: Toward GANs Exploiting Multiple Samples. arXiv preprint arXiv:2304.12098. 2023 April 24.
[2] Kang M, Shin J, Park J. Studiogan: A taxonomy and benchmark of gans for image synthesis. arXiv preprint arXiv:2206.09479. 2022 Jun 19.