Keras GAN library

Introduction

Implementation of GAN papers, all using cifar10 dataset in this project.

ATTENTION: To compare the differences of GAN methods, the hyperparameters in this project are not exactly same as papers. Architecture of generators and discriminators are as similar as possible, and using same optimizer setting.

Environment

python==3.6
tensorflow==1.4.0
keras==2.1.0

Implemented Papers

• DCGAN - Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks link
• LSGAN - Loss-Sensitive Generative Adversarial Networks on Lipschitz Densities link
• WGAN-GP - Improved Training of Wasserstein GANs link
• SNGAN - Spectral Normalization for Generative Adversarial Networks link

Results

Name	50 epochs
DCGAN
LSGAN
WGAN-GP
SNGAN