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.
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)