This is a PyTorch implementation of a new prior ("Variational Mixture of Posteriors" prior, or VampPrior for short) for the variational auto-encoder framework with one layer and two layers of stochastic hidden units as described in the following paper:

• Jakub M. Tomczak, Max Welling, VAE with a VampPrior, arXiv preprint, 2017

The code is compatible with:

• pytorch 0.2.0

The experiments can be run on the following datasets:

• static MNIST: links to the datasets can found at link;
• binary MNIST: the dataset is loaded from PyTorch;
• OMNIGLOT: the dataset could be downloaded from link;
• Caltech 101 Silhouettes: the dataset could be downloaded from link.
• Frey Faces: the dataset could be downloaded from link.
• Histopathology Gray: the dataset could be downloaded from link;
• CIFAR 10: the dataset is loaded from PyTorch.

1. Set-up your experiment in experiment.py.
2. Run experiment:

python experiment.py

You can run a vanilla VAE, a one-layered VAE or a two-layered VAE with the standard prior or the VampPrior by setting model_name argument to either: (i) vae or vae_2level for MLP, (ii) convvae or convvae_2level for convnets, and specifying prior argument to either standard or vampprior.

Please cite our paper if you use this code in your research:

@article{TW:2017,
  title={{VAE with a VampPrior}},
  author={Tomczak, Jakub M and Welling, Max},
  journal={arXiv},
  year={2017}
}

• CIFAR 10 is not working properly (NaN after some epochs).

The research conducted by Jakub M. Tomczak was funded by the European Commission within the Marie Skłodowska-Curie Individual Fellowship (Grant No. 702666, ”Deep learning and Bayesian inference for medical imaging”).