Installation

Installation#

Note

In the following, commands to be run in the shell are displayed here with a leading $. You do not have to type it.

Commands starting with python may need to be run with python3 instead, depending on your system.

Pre-requisites#

GPry requires Python (version ≥ 3.9, check the output of $ python --version) and an up to date version of the Python package manager pip (to install: $ python -m ensurepip; to update: $ python -m pip install pip --upgrade). Most of the requisites will automatically be installed by pip.

Installing GPry#

To install GPry or upgrade it to the latest release, simply type in a terminal

$ python -m pip install gpry --upgrade

This should install GPry and all dependencies.

Of course you can also fork/clone the repo or download the source code and install that. Be aware though, that we are actively developing the code in git, though most of the development occurs in branches.

Configuring and using MPI#

Though optional, in order to make GPry as fast as possible, you will need a working MPI installation and the necessary Python bindings.

In order to install MPI in a Debian-like system, do:

$ sudo apt install openmpi-bin libopenmpi-dev

To get the MPI bindings, install the mpi4py package with pip.

If successful, you should be able to run the following command and get MPI is working as an output:

$ mpirun -n 2 python -c "exec('from mpi4py import MPI\nif MPI.COMM_WORLD.Get_rank() == 1: print(\"MPI is working\")')"

Once MPI is working, you can simply run your Python script as

$ mpirun -n 2 python your_script.py

Note

GPry takes care internally of handling parallelization in the parts of the code that are supposed to be run by a single process, e.g. the plotting scripts, so that they do not produce multiple outputs/plots.

Installing Nested Samplers#

In order to use the highly parallelizable and more exploratory NORA acquisition engine, you will need to install one of the following nested samplers:

  • PolyChord: it is the preferred option, since in combination with MPI it is very fast. To install it, follow the instructions at PolyChord/PolyChordLite. Try the MPI example mentioned there to make sure it works.

  • UltraNest: it is the default if PolyChord is not present. It is slower but easier to install: simply pip install ultranest. It does take advantage of MPI parallelization.

  • nessai: [EXPERIMENTAL SUPPORT] this sampler uses ML to increase efficiency in posteriors with non-linear degeneracies, but cannot take advantage of MPI parallelization. To install it, follow the instructions at https://nessai.readthedocs.io/en/latest/installation.html.

To check that any of the three have been correctly installed, check that the following command does not throw an error: $ python -c "import SAMPLER", replacing SAMPLER with pypolychord, ultranest, nessai.

Contents