/WGAN-GP-tensorflow

Tensorflow Implementation of Paper "Improved Training of Wasserstein GANs"

Primary LanguagePythonMIT LicenseMIT

WGAN-GP-tensorflow

Tensorflow implementation of paper "Improved Training of Wasserstein GANs".

gif

  • 0 epoch

epoch0

  • 25 epoch

img

  • 50 epoch

epoch50

  • 100 epoch

img

  • 150 epoch

img

Prerequisites

  • Python 2.7 or 3.5
  • Tensorflow 1.3+
  • SciPy
  • Aligned&Cropped celebA dataset(download)
  • (Optional) moviepy (for visualization)

Usage

  • Download aligned&cropped celebA dataset(link) and unzip at ./data/img_align_celeba

  • Train:

    $ python main.py --train

    Or you can set some arguments like:

    $ python main.py --dataset=celebA --max_epoch=50 --learning_rate=1e-4 --train

  • Test:

    $ python main.py

Acknowledge

Based on the implementation carpedm20/DCGAN-tensorflow, LynnHo/DCGAN-LSGAN-WGAN-WGAN-GP-Tensorflow, shekkizh/WassersteinGAN.tensorflow and igul222/improved_wgan_training.