Wassertein Generative Adversarial Network(WGAN)

This is simple demo of Wassertein GAN on Anime dataset. The model needs around 20000 iterations to get reasonable result. The running time could be around 3 hours on GTX 1070 GPU.

How to start

$python main.py

Data

Anime Data was crawled from Japanese Anime Websites randomly and preprocessed through anime face detection code, for example lbpcascade. I collected around 20000 images as training data. Please do not ask me for sharing the data due to potential copyright issue.

Stability

Before using Wassertein GAN, I tried multiple hacking technology such as mini-batch discrimination and tricks. But, unfortunately, none of them really works in general but really depends on how lucky you are.

Demo