Variational Inference by Policy Search

Variational Inference by Policy Search (VIPS) is a method for learning Gaussian Mixture Model approximations of intractable probability density functions for the purpose of inference (e.g. sampling).
VIPS does not require knowledge about gradients or normalizing constants. The optimization leverages insights from Policy Search (hence the name) by using information-geometric trust region to improve the approximation in a controlled manner for better stability and exploration.
More details about the method can be found in the paper. For API documentation refer to doc/html/index.html.

Installation

This implementation can be used with Python 3. However, most of the algorithm is implemented in C++ and interfaced via SWIG.
It should actually be easy to use this as a pure C++ library, however the learning loop is only implemented in python for now (see python/sampler/VIPS/VIPS.py).
Installation consists of two parts. You first need to install some libraries needed by VIPS (OpenBlas, Armadillo and NLOPT). Afterwards the C++ part of VIPS can be compiled.

Compiling the C++ Libraries

The following instructions will install nlopt and armadillo+OpenBLAS locally into ~/libs. When installing into a different directory, the CMakeList.txt needs to be modified.
Instructions are based on https://gist.github.com/BERENZ/ff274ebbf00ee111c708.

Install OpenBlas

In some temporary directory run:

cd /tmp
git clone git://github.com/xianyi/OpenBLAS.git
cd OpenBLAS/
git checkout develop
mkdir ~/libs
USE_OPENMP=1 NO_SHARED=1 COMMON_OPT=" -O2 -march=native" make
make PREFIX=~/libs/OpenBlas NO_SHARED=1 install

Install armadillo

Download armadillo into ~/libs

cd ~/libs
wget http://sourceforge.net/projects/arma/files/armadillo-9.100.5.tar.xz

You can also check for a newer version at http://arma.sourceforge.net/download.html

Unpack and configure it (Don't make it, we just need the headers)

tar xJvf armadillo-9.100.5.tar.xz
mv armadillo-9.100.5 armadillo
rm armadillo-9.100.5.tar.xz
cd armadillo
./configure

Install NLOPT

wget http://ab-initio.mit.edu/nlopt/nlopt-2.4.2.tar.gz
tar xzvf nlopt-2.4.2.tar.gz
cd nlopt-2.4.2/
./configure --enable-shared --prefix=$HOME/libs/nlopt && make && make install

add the following lines to your ~/.bashrc:

LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/libs/nlopt/lib
export LD_LIBRARY_PATH

Compiling VIPS

Make sure that swig is installed

sudo apt install swig

Change into ./c++ and compile the project:

mkdir build
cd build
cmake ..
make install -j4

Getting Started

Scripts for the experiments of the paper can be found in python/experiments/VIPS/. Make sure that python is in your path when running the scripts or run them as modules, e.g.

/VIPS/python/$ python3 -m experiments.VIPS.breast_cancer

Check python/experiments/VIPS/rosenbrock.py for an easy example of how to use VIPS for approximating a given density function.

Citation

If you use this package in your own work please cite our paper:

Arenz, O.; Zhong, M.; Neumann, G. Efficient Gradient-Free Variational Inference using Policy Search. Proceedings of the 35th International Conference on Machine Learning. 2018.

License

This project is licensed under the MIT License - see the LICENSE file for details.