Triple Generative Adversarial Nets (Triple-GAN)

Chongxuan Li, Kun Xu, Jun Zhu and Bo Zhang

Code for reproducing most of the results in the paper. Triple-GAN: a unified GAN model for classification and class-conditional generation in semi-supervised learning.

Warning: the code is still under development.

Triple-GAN-V2 and code in Pytorch!

We propose Triple-GAN-V2 built upon mean teacher classifier and projection discriminator with spectral norm and implement Triple-GAN in Pytorch. See the source code at taufikxu/Triple-GAN

Envoronment settings and libs we used in our experiments

This project is tested under the following environment setting.

• OS: Ubuntu 16.04.3
• GPU: Geforce 1080 Ti or Titan X(Pascal or Maxwell)
• Cuda: 8.0, Cudnn: v5.1 or v7.03
• Python: 2.7.14(setup with Miniconda2)
• Theano: 0.9.0.dev-c697eeab84e5b8a74908da654b66ec9eca4f1291
• Lasagne: 0.2.dev1
• Parmesan: 0.1.dev1

Python Numpy Scipy Theano Lasagne(version 0.2.dev1) Parmesan

Thank the authors of these libs. We also thank the authors of Improved-GAN and Temporal Ensemble for providing their code. Our code is widely adapted from their repositories.

Results

Triple-GAN can achieve excellent classification results on MNIST, SVHN and CIFAR10 datasets, see the paper for a comparison with the previous state-of-the-art. See generated images as follows:

Comparing Triple-GAN (right) with GAN trained with feature matching (left)

Generating images in four specific classes (airplane, automobile, bird, horse)

Disentangling styles from classes (left: data, right: Triple-GAN)

Class-conditional linear interpolation on latent space