Wasserstein GAN with gradient penalty

Keras model and tensorflow optimization of "improved Training of Wasserstein GANs"

please download image training data before learning.

python main.py --datadir (path to the directory contains images)

sample generated image (using lsun dataset, church_outdoor category)

! it generate well-seems fake images, but the loss (both of genertor and of discriminator) getting larger along learning.

! so it may have some imcomplete points, want someone to check it, thanks.