DPPy: Sampling Determinantal Point Processes with Python

"Anything that can go wrong, will go wrong" − Murphy's Law

Introduction

Determinantal point processes (DPPs) are specific probability distributions over clouds of points that have been popular as models or computational tools across physics, probability, statistics, and more recently of booming interest in machine learning. Sampling from DPPs is a nontrivial matter, and many approaches have been proposed. DPPy is a Python library that puts together all exact and approximate sampling algorithms for DPPs.

Installation

DPPy works with Python 3.6+.

Dependencies

This project depends on the following libraries, which are automatically downloaded during installation:

• NumPy
• SciPy
• Matplotlib

The following dependencies are optional, and unlock extra functionality if installed:

• Networkx to play with uniform spanning trees .

• CVXOPT to use the zono_sampling MCMC sampler for finite DPPs. CVXOPT itself requires GCC,

  • On MAC it comes with Xcode

  • On UNIX, use your package manager (apt, yum etc)

    sudo apt install -qq gcc g++

• Sphinx to modify and rebuild the documentation

Installation instructions

DPPy is now available on PyPI

pip install dppy

For a full installation, including extra dependencies and dependencies necessary to build the documentation (see above), use pip install dppy['zonotope','trees','docs'].

Note that only stable DPPy releases are available on PyPI, and recently included improvements might only be available on the git master branch. If you want to work with the latest version of DPPy

1. If you have a GitHub account please consider forking DPPy and use git to clone your copy of the repo

   cd <directory_of_your_choice>
   git clone https://github.com/<username>/DPPy.git

2. If you only use git, clone this repository

   cd <directory_of_your_choice>
   git clone https://github.com/guilgautier/DPPy.git

3. Otherwise simply download the project

4. In any case, install the project with

   cd DPPy
   pip install .

Use pip install .['zonotope','trees','docs'] to perform a full install from a local source tree.

How to use it

The main DPPy documentation is available online at http://dppy.readthedocs.io. There are also some interactive tutorials using Jupyter available at https://github.com/guilgautier/DPPy/blob/master/notebooks/. For more details, check below.

Tutorials in Jupyter notebooks

You can read and work on these interactive tutorial Notebooks, directly from your web browser, without having to download or install Python or anything. Just click, wait a little bit, and play with the notebook!

Building the documentation

The documentation is generated locally with Sphinx and then built online by ReadTheDocs. If you wish to contribute to the documentation or just play with it locally, you can install the necessary dependencies and then:

• Generate the docs locally

  cd DPPy/docs
  make html

• Open the local HTML version of the documentation located at DPPy/docs/_build/html/index.html

  open _build/html/index.html

How to cite this work?

We wrote a companion paper to DPPy which got accepted for publication in the Machine Learning Open Source Software track of JMLR.

If you use the DPPy toolbox, please consider citing it with this piece of BibTeX:

@article{GPBV19,
  author = {Gautier, Guillaume and Polito, Guillermo and Bardenet, R{\'{e}}mi and Valko, Michal},
  journal = {Journal of Machine Learning Research - Machine Learning Open Source Software (JMLR-MLOSS)},
  title = {{DPPy: DPP Sampling with Python}},
  keywords = {Computer Science - Machine Learning, Computer Science - Mathematical Software, Statistics - Machine Learning},
  url = {http://jmlr.org/papers/v20/19-179.html},
  year = {2019},
  archivePrefix = {arXiv},
  arxivId = {1809.07258},
  note = {Code at http://github.com/guilgautier/DPPy/ Documentation at http://dppy.readthedocs.io/}
}

Many of the algorithms implemented in DPPy also have associated literature that you should consider citing. Please refer to the bibliography on the documentation page and docstrings of each specific DPP sampler for details.

Reproducibility

We would like to thank Guillermo Polito for leading our reproducible research workgroup, this project owes him a lot.

Take a look at the corresponding booklet to learn more on how to make your research reproducible!