start0821/pytorch-gan-collections

PyTorch implementation of DCGAN, WGAN-GP and SNGAN.

Collections of GANs

Pytorch implementation of unsupervised GANs.

For more defails about calculating Inception Score and FID using pytorch can be found in pytorch-inception-score-fid

Models

• DCGAN
• WGAN
• WGAN-GP
• SN-GAN

Requirements

• Initialize metric submoudle

  git submodule update --init
  

• Install python packages

  pip install -U pip setuptools
  pip install -r requirements.txt

Results

Model	Dataset	Inception Score	FID
DCGAN	CIFAR10	6.04(0.05)	47.90
WGAN(CNN)	CIFAR10	6.64(0.6)	33.27
WGAN-GP(CNN)	CIFAR10	7.47(0.06)	24.00
WGAN-GP(ResNet)	CIFAR10	7.74(0.10)	21.89
SNGAN(CNN)	CIFAR10	7.44(0.11)	24.94
SNGAN(ResNet)	CIFAR10	8.22(0.13)	16.24

Examples

• DCGAN

  

• WGAN(CNN)

  

• WGAN-GP(CNN)

  

• WGAN-GP(ResNet)

  

• SNGAN(CNN)

  

• SNGAN(ResNet)

  

Reproduce

• Download cifar10.test.npz for calculating FID score. Then, create folder stats for the npz files

  stats
  ├── cifar10.test.npz
  ├── cifar10.train.npz
  └── stl10.unlabeled.48.npz
  

• Train from scratch

  # DCGAN
  python dcgan.py --flagfile ./config/DCGAN_CIFAR10.txt
  # WGAN(CNN)
  python wgan.py --flagfile ./config/WGAN_CIFAR10_CNN.txt
  # WGAN-GP(CNN)
  python wgangp.py --flagfile ./config/WGANGP_CIFAR10_CNN.txt
  # WGAN-GP(ResNet)
  python wgangp.py --flagfile ./config/WGANGP_CIFAR10_RES.txt
  # SNGAN(CNN)
  python sngan.py --flagfile ./config/SNGAN_CIFAR10_CNN.txt
  # SNGAN(ResNet)
  python sngan.py --flagfile ./config/SNGAN_CIFAR10_RES.txt

  Though the training procedures of different GANs are almost identical, I still separate different methods into different files for clear reading.

Learning curve

Change Log

• 2021-04-16
  • Update pytorch to 1.8.1
  • Move metrics to submodule.
  • Evaluate FID on CIFAR10 test set instead of training set.
  • Fix cifar10.test.npz download link and sample images.