/getdist

MCMC sample analysis, kernel densities, plotting, and GUI

Primary LanguagePythonOtherNOASSERTION

GetDist

GetDist:MCMC sample analysis, plotting and GUI
Author: Antony Lewis
Homepage:https://getdist.readthedocs.io
Source:https://github.com/cmbant/getdist
Reference:https://arxiv.org/abs/1910.13970
https://travis-ci.org/cmbant/getdist.svg?branch=master https://img.shields.io/pypi/v/GetDist.svg?style=flat https://readthedocs.org/projects/getdist/badge/?version=latest https://img.shields.io/badge/arXiv-1910.13970-b31b1b.svg?color=0B6523

Description

GetDist is a Python package for analysing Monte Carlo samples, including correlated samples from Markov Chain Monte Carlo (MCMC).

  • Point and click GUI - select chain files, view plots, marginalized constraints, LaTeX tables and more
  • Plotting library - make custom publication-ready 1D, 2D, 3D-scatter, triangle and other plots
  • Named parameters - simple handling of many parameters using parameter names, including LaTeX labels and prior bounds
  • Optimized Kernel Density Estimation - automated optimal bandwidth choice for 1D and 2D densities (Botev et al. Improved Sheather-Jones method), with boundary and bias correction
  • Convergence diagnostics - including correlation length and diagonalized Gelman-Rubin statistics
  • LaTeX tables for marginalized 1D constraints

See the Plot Gallery and tutorial (run online) and GetDist Documentation.

Getting Started

Install getdist using pip:

$ pip install getdist

or from source files using:

$ python setup.py install

or:

$ pip install -e /path/to/source/

You can test if things are working using the unit test by running:

$ python -m unittest getdist.tests.getdist_test

Check the dependencies listed in the next section are installed. You can then use the getdist module from your scripts, or use the GetDist GUI (getdist-gui command).

Once installed, the best way to get up to speed is probably to read through the Plot Gallery and tutorial.

Dependencies

  • Python 3.6+
  • matplotlib 2.2+ (3.1+ recommended)
  • scipy
  • PySide2 - optional, only needed for GUI
  • Working LaTeX installation (not essential, only for some plotting/table functions)

Python distributions like Anaconda have most of what you need (except for LaTeX).

To use the GUI you need PySide2. See the GUI docs for suggestions on how to install.

Algorithm details

Details of kernel density estimation (KDE) algorithms and references are give in the GetDist notes arXiv:1910.13970.

Samples file format

The GetDist GUI (and getdist.loadMCSamples function) read parameter sample/chain files in plain text format. In general there are a set of plain text files of the form:

xxx_1.txt
xxx_2.txt
...
xxx.paramnames
xxx.ranges

where "xxx" is some root file name.

The .txt files are separate chain files (there can also be just one xxx.txt file). Each row of each sample .txt file is in the format

weight like param1 param2 param3 ...

The weight gives the number of samples (or importance weight) with these parameters. like gives -log(likelihood), and param1, param2... are the values of the parameters at the sample point. The first two columns can be 1 and 0 if they are not known or used.

The .paramnames file lists the names of the parameters, one per line, optionally followed by a LaTeX label. Names cannot include spaces, and if they end in "*" they are interpreted as derived (rather than MCMC) parameters, e.g.:

x1   x_1
y1   y_1
x2   x_2
xy*  x_1+y_1

The .ranges file gives hard bounds for the parameters, e.g.:

x1  -5 5
x2   0 N

Note that not all parameters need to be specified, and "N" can be used to denote that a particular upper or lower limit is unbounded. The ranges are used to determine densities and plot bounds if there are samples near the boundary; if there are no samples anywhere near the boundary the ranges have no affect on plot bounds, which are chosen appropriately for the range of the samples.

There can also optionally be a .properties.ini file, which can specify burn_removed=T to ensure no burn in is removed, or ignore_rows=x to ignore the first fraction x of the file rows (or if x > 1, the specified number of rows).

Loading samples

To load an MCSamples object from text files do:

from getdist import loadMCSamples
samples = loadMCSamples('/path/to/xxx', settings={'ignore_rows':0.3})

Here settings gives optional parameter settings for the analysis. ignore_rows is useful for MCMC chains where you want to discard some fraction from the start of each chain as burn in (use a number >1 to discard a fixed number of sample lines rather than a fraction). The MCSamples object can be passed to plot functions, or used to get many results. For example, to plot marginalized parameter densities for parameter names x1 and x2:

from getdist import plots
g = plots.get_single_plotter()
g.plot_2d(samples, ['x1', 'x2'])

When you have many different chain files in the same directory, plotting can work directly with the root file names. For example to compare x and y constraints from two chains with root names xxx and yyy:

from getdist import plots
g = plots.get_single_plotter(chain_dir='/path/to/', analysis_settings={'ignore_rows':0.3})
g.plot_2d(['xxx','yyy'], ['x', 'y'])

MCSamples objects can also be constructed directly from numpy arrays in memory, see the example in the Plot Gallery.

GetDist script

If you have chain files on on disk, you can also quickly calculate convergence and marginalized statistics using the getdist script:

usage: getdist [-h] [--ignore_rows IGNORE_ROWS] [-V] [ini_file] [chain_root]

GetDist sample analyser

positional arguments:

ini_file .ini file with analysis settings (optional, if omitted uses defaults

chain_root Root name of chain to analyse (e.g. chains/test), required unless file_root specified in ini_file

optional arguments:
-h, --help show this help message and exit
--ignore_rows IGNORE_ROWS
 set initial fraction of chains to cut as burn in (fraction of total rows, or >1 number of rows); overrides any value in ini_file if set
--make_param_file MAKE_PARAM_FILE
 Produce a sample distparams.ini file that you can edit and use when running GetDist
-V, --version show program's version number and exit

where ini_file is optionally a .ini file listing key=value parameter option values, and chain_root is the root file name of the chains. For example:

getdist distparams.ini chains/test_chain

This produces a set of files containing parameter means and limits (.margestats), N-D likelihood contour boundaries and best-fit sample (.likestats), convergence diagnostics (.converge), parameter covariance and correlation (.covmat and .corr), and optionally various simple plotting scripts. If no ini_file is given, default settings are used. The ignore_rows option allows some of the start of each chain file to be removed as burn in.

To customize settings you can run:

getdist --make_param_file distparams.ini

to produce the setting file distparams.ini, edit it, then run with your custom settings.

GetDist GUI

Run getdist-gui to run the graphical user interface. This requires PySide2, but will run on Windows, Linux and Mac. It allows you to open a folder of chain files, then easily select, open, plot and compare, as well as viewing standard GetDist outputs and tables. See the GUI Readme.

Using with CosmoMC and Cobaya

This GetDist package is general, but is mainly developed for analysing chains from the CosmoMC and Cobaya sampling programs. No need to install this package separately if you have a full CosmoMC installation; the Cobaya installation will also install GetDist as a dependency. Detailed help is available for plotting Planck chains and using CosmoMC parameter grids in the Readme.

Citation

You can refer to the notes:

@article{Lewis:2019xzd,
 author         = "Lewis, Antony",
 title          = "{GetDist: a Python package for analysing Monte Carlo
                   samples}",
 year           = "2019",
 eprint         = "1910.13970",
 archivePrefix  = "arXiv",
 primaryClass   = "astro-ph.IM",
 SLACcitation   = "%%CITATION = ARXIV:1910.13970;%%",
 url            = "https://getdist.readthedocs.io"
}

and references therein as appropriate.