Awesome

Good Semi-Supervised Learning that Requires a Bad GAN

This is the code we used in our paper

Good Semi-supervised Learning that Requires a Bad GAN

Zihang Dai*, Zhilin Yang*, Fan Yang, William W. Cohen, Ruslan Salakhutdinov (*: equal contribution)

NIPS 2017

Requirements

The repo supports python 2.7 + pytorch 0.1.12. To install pytorch 0.1.12, run conda install pytorch=0.1.12 cuda80 -c soumith.

Get Pretrained PixelCNN Model

mkdir model
cd model
wget http://kimi.ml.cmu.edu/mnist.True.3.best.pixel

Run the Code

To reproduce our results on MNIST

python mnist_trainer.py

To reproduce our results on SVHN

python svhn_trainer.py

To reproduce our results on CIFAR-10

python cifar_trainer.py

Results

Here is a comparison of different models using standard architectures without ensembles (100 labels on MNIST, 1000 labels on SVHN, and 4000 labels on CIFAR):