py-coda

Python bindings for Monte Carlo Markov chain (MCMC) diagnostics using the coda package in R

Installation:

1. Install R (https://www.r-project.org/) and rpy2

On Mac with Homebrew

brew tap homebrew/science
brew install r
pip install rpy2

or Ubuntu

sudo apt-get install r-base r-base-dev python-rpy2

2. Install the coda package

mkdir $HOME/Rpackages
echo R_LIBS_USER="$HOME/Rpackages" > ~/.Renviron
Rscript -e 'install.packages("coda", repos="http://cran.us.r-project.org")'

3. Install py-coda

git clone https://github.com/surhudm/py-coda.git
cd py-coda
python setup.py install

Run MCMC diagnostics:

Prepare MCMC file say chainfile.out where every line is the parameter set at every step of the chain. Prepare another file where each line lists the parameter name, write this parameter name in latex form, as it will get set as a label on the plots. For examples, see the files chainfile.ind and chainfile.out (the latter file only contains 100 lines to prevent the repository from being bulky). Then you could run the diagnostics as follows:

from py_coda import read_pycoda
mcmcobj = read_pycoda("chainfile.out", "chainnew.ind", thin=1)
mcmcobj.geweke()
mcmcobj.get_stats()
mcmcobj.plot_traces()
mcmcobj.plot_autocorr()
mcmcobj.heidelberger_welch()

In your python code you could even use the python bindings for coda directly. For example, if you have a numpy array with shape (Nchain, Nparam) which stores the MCMC where Nchain is the number of iterations of the MCMC and Nparam is the number of parameters, and an array of labels for your parameters, you could simply do

from py_coda import mcmc
mcmcobj = mcmc(labels, data, thin=1)
mcmcobj.geweke()
mcmcobj.get_stats()
mcmcobj.plot_traces()
mcmcobj.plot_autocorr()
mcmcobj.heidelberger_welch()

Please file bug reports or issues if you find any.