Adversarial Images for Variational Autoencoders

To be presented at the Adversarial Training Workshop at NIPS 2016, Barcelona.

Arxiv link will be posted in the near future!

Please cite our work:

Pedro Tabacof, Julia Tavares, and Eduardo Valle. Adversarial Images for Variational Autoencoders. Adversarial Training Workshop, NIPS. 2016.

Requirements

Theano

Lasagne

Parmesan

Notebooks

To reproduce our experiments, simply run the notebooks.

There are some options that can be readily changed, the most important one being do_train_model: Set it to True to train the model from scratch, or to False to use the pretrained models in the params folder.

Files

Experiments

adv: Adversarial images for (variational) autoencoders

clf: Adversarial images for classifier experiments

Architectures

ae: Deterministic autoencoders

vae: Variational autoencoders

Datasets

mnist: MNIST dataset

svhn: SVHN dataset

Pretrained weights

params: Pretrained AEs, VAEs and CLFs

Results

results: folder with CSVs containing the experiments results -- to be used for plotting

results.ipynb: Plotting results from the folder above