takerum / vat

Code for reproducing the results on the MNIST dataset in the paper "Distributional Smoothing with Virtual Adversarial Training"

Virtual Adversarial training (VAT) implemented with Theano

Python code for reproducing the results showed in the paper:"Distributional Smoothing with Virtual Adversarial Training" http://arxiv.org/abs/1507.00677

Required libraries

python 2.7, numpy 1.9, theano 0.7.0, docopt 0.6.2

Examples on synthetic dataset

Model's contours on synthetic datasets with different regularization methods (Fig.3,4 in our paper)

./vis_model_contours.sh

The coutour images will be saved in ./figure.

Examples on MNIST dataset

Download mnist.pkl

cd dataset
./download_mnist.sh

###VAT for supervised learning on MNIST dataset

python train_mnist_sup.py --cost_type=VAT_finite_diff --epsilon=2.1 --layer_sizes=784-1200-600-300-150-10 --save_filename=<filename>

###VAT for semi-supervised learning on MNIST dataset (with 100 labeled samples)

python train_mnist_semisup.py --cost_type=VAT_finite_diff --epsilon=0.3 --layer_sizes=784-1200-1200-10 --num_labeled_samples=100 --save_filename=<filename>

After finish training, the trained classifer will be saved with <filename> in ./trained_model.

You can obtain a test error of the trained classifier saved with <filename> by the following command:

python test_mnist.py --load_filename=<filename>

.

If you find bug or problem, please report it!