Collections of GANs

Pytorch implementation of basic unsupervised GANs on CIFAR10.

For more defails about calculating Inception Score and FID using pytorch can be found here pytorch_gan_metrics.

Models

• DCGAN
• WGAN
• WGAN-GP
• SN-GAN

Requirements

• Install python packages

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

Results

The FID is calculated by 50k generated images and CIFAR10 train set.

Examples

• DCGAN

  

• WGAN(CNN)

  

• WGAN-GP(CNN)

  

• WGAN-GP(ResNet)

  

• SNGAN(CNN)

  

• SNGAN(ResNet)

  

Reproduce

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

  stats
  └── cifar10.train.npz
  

• Train from scratch

  Different methods are separated into different files for clear reading.

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

Learning Curves

Change Log

• 2022-01-10

  • Update pytorch to 1.10.1 and CUDA 11.3
  • Use pytorch_gan_metrics to calculate FID and Inception Score
  • Use 50k generated images and CIFAR10 train set to calculate FID
  • Fix default parameters especially for wgan.py

• 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.