Sylvester normalizing flows for variational inference

Pytorch implementation of Sylvester normalizing flows, based on our paper:

Sylvester normalizing flows for variational inference (UAI 2018)
Rianne van den Berg*, Leonard Hasenclever*, Jakub Tomczak, Max Welling

*Equal contribution

Requirements

The latest release of the code is compatible with:

• pytorch 1.0.0

• python 3.7

Thanks to Martin Engelcke for adapting the code to provide this compatibility.

Version v0.3.0_2.7 is compatible with:

• pytorch 0.3.0 WARNING: More recent versions of pytorch have different default flags for the binary cross entropy loss module: nn.BCELoss(). You have to adapt the appropriate flags if you want to port this code to a later vers
  ion.

• python 2.7

Data

The experiments can be run on the following datasets:

• static MNIST: dataset is in data folder;
• OMNIGLOT: the dataset can be downloaded from link;
• Caltech 101 Silhouettes: the dataset can be downloaded from link.
• Frey Faces: the dataset can be downloaded from link.

Usage

Below, example commands are given for running experiments on static MNIST with different types of Sylvester normalizing flows, for 4 flows:

Orthogonal Sylvester flows
This example uses a bottleneck of size 8 (Q has 8 columns containing orthonormal vectors).

python main_experiment.py -d mnist -nf 4 --flow orthogonal --num_ortho_vecs 8 

Householder Sylvester flows
This example uses 8 Householder reflections per orthogonal matrix Q.

python main_experiment.py -d mnist -nf 4 --flow householder --num_householder 8

Triangular Sylvester flows

python main_experiment.py -d mnist -nf 4 --flow triangular 

To run an experiment with other types of normalizing flows or just with a factorized Gaussian posterior, see below.

Factorized Gaussian posterior

python main_experiment.py -d mnist --flow no_flow

Planar flows

python main_experiment.py -d mnist -nf 4 --flow planar

Inverse Autoregressive flows
This examples uses MADEs with 320 hidden units.

python main_experiment.py -d mnist -nf 4 --flow iaf --made_h_size 320

More information about additional argument options can be found by running ```python main_experiment.py -h```

Cite

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

@inproceedings{vdberg2018sylvester,
  title={Sylvester normalizing flows for variational inference},
  author={van den Berg, Rianne and Hasenclever, Leonard and Tomczak, Jakub and Welling, Max},
  booktitle={proceedings of the Conference on Uncertainty in Artificial Intelligence (UAI)},
  year={2018}
}