Skip to content

Latest commit

 

History

History
39 lines (29 loc) · 2.2 KB

README.md

File metadata and controls

39 lines (29 loc) · 2.2 KB
Raw

VAE with a Variational Mixture of Posteriors "VampPrior"

This is an implementation of the following paper in Tensorflow 2.0:

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

We hereby compare the performance 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.

Models

We provide a vanilla version of a VAE and a Hierarchical one with two level priors. Each architecture can be used with and without the VampPrior. The VampPrior let you train the pseudo-inputs or randomly choose them from the data. The models training can be monitored with both training and validation losses. The overall performance is quantified by the marginal test log likelihood. You can run a vanilla VAE, a two-layered HVAE with the standard prior or the VampPrior by setting architecture argument to either: (i) runner.Architecture.HVAE or architecture=runner.Architecture.VANILLA for HVAE, (ii) and specifying prior_configuration argument to either runner.PriorConfiguration.SG or runner.PriorConfiguration.VAMPGEN or PriorConfiguration.VAMPDATA.

Requirements

The code was implemeted with:

  • Tensorflow >= 2.0
  • Tensorflow Probability

Data

The experiments were conducted on the following datasets:

  • static MNIST: links to the datasets can found at link
  • OMNIGLOT: the dataset could be downloaded from link
  • Caltech 101 Silhouettes: the dataset could be downloaded from link

You can specify the dataset_key argument to either runner.Dataset.CALTECH or runner.Dataset.MNIST or runner.Dataset.OMNIGLOT

Set an Experiment

We provide a test framework for running the different experiments, refer to experiments/test_runner.py for a quick tutorial.

References

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