The original is available at https://github.com/pfnet-research/sngan_projection.

This is unofficial PyTorch implementation of sngan_projection.
This does not reproduce the experiments and results reported in the paper due to the lack of GPUs.
This repository does some experiments on images of size 64x64.

Some results are on issues with results label.

Spectral Normalization for Generative Adversarial Networks
Takeru Miyato, Toshiki Kataoka, Masanori Koyama, Yuichi Yoshida
OpenReview: https://openreview.net/forum?id=B1QRgziT-
arXiv: https://arxiv.org/abs/1802.05957

cGANs with Projection Discriminator
Takeru Miyato, Masanori Koyama
OpenReview: https://openreview.net/forum?id=ByS1VpgRZ
arXiv: https://arxiv.org/abs/1802.05637

• Python 3.6.4
• PyTorch 0.4.1
• torchvision 0.2.1
• NumPy: Used in FID score calculation and data loader
• SciPy: Used in FID score calculation
• tensorflow (optional)
• tensorboardX (optional)
• tqdm: Progressbar and Log

If you want to use tensorboard for beautiful training update visualization, please install tensorflow and tensorboardX.
When using only tensorboard, tensorflow cpu is enough.

Dockerfiles for pytorch 1.0 environment and tensorboard are added. PyTorch 1.0 Dockerfile requires an nvidia driver that supports CUDA 9.2. Also, this dockerized environment needs some environment variables:

• DATA: Path to dataset
• RESULTS: Path to save results
• PORT: Port number for jupyter notebook.

• tiny ImageNet¹.

Tiny Imagenet has 200 classes. Each class has 500 training images, 50 validation images, and 50 test images.

Default parameters are the same as the original Chainer implementation.

• to train cGAN with projection discriminator: run train_64.py with --cGAN option.
• to train cGAN with concat discriminator: run train_64.py with both --cGAN and --dis_arch_concat.
• to run without tensorboard, please add --no_tensorboard.
• to calculate FID, add --calc_FID (not tested)
• to use make discriminator relativistic, add --relativistic_loss or -relloss (not tested)

To see all the available arguments, run python train_64.py --help.

• implement super-resolution (cGAN)

1. https://github.com/pfnet-research/sngan_projection
2. https://github.com/mseitzer/pytorch-fid: FID score
3. https://github.com/naoto0804/pytorch-AdaIN: Infinite Sampler of DataLoader