/pytorch.sngan_projection

An unofficial PyTorch implementation of SNGAN (ICLR 2018) and cGANs with projection discriminator (ICLR 2018).

Primary LanguagePythonMIT LicenseMIT

UPDATE 2019/07/14: C-GAN has a bug in network definition!
Thank you @UdonDa for reporting and pointing out (issue#28)

UPDATE 2019/05/24: Current implementation of FID score is incorrect!
Thank you @youngjung for the report & fix suggestions (issue#25)!

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

SNGAN and cGANs with projection discriminator

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.

SNGAN

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

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

Requirements

  • 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.

Docker environment

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.

Dataset

  • tiny ImageNet1.

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

Training configuration

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.

TODO

  • implement super-resolution (cGAN)

Acknowledgement

  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

Footnotes

  1. https://tiny-imagenet.herokuapp.com/ ↩