pyEPABC

A Python implementation of EP-ABC for likelihood-free, probabilistic inference.

Author:	Sebastian Bitzer
Contact:	[email protected]

This code is based on Simon Barthelmé's Matlab implementation of EP-ABC as described in the corresponding article. See the documentation of run_EPABC for further details.

This implementation requires Python 3.

Source Installation

Clone the repository and use the Python setuptools to install pyEPABC, e.g. in the created directory call:

python setup.py install

For the DDM-example below to work, installation is a bit more complicated, see the example for further details.

Examples

The examples-folder contains two examples demonstrating the usage of pyEPABC: testGauss and testDDM. Instructions for how to run these examples follow.

Simple Gaussian example

This example compares EPABC to an analytic solution for the simplest case you can imagine: Estimating the mean of some data points. The simplicity of the problem allows that the posterior distribution and log marginal likelihood can be computed analytically so that the EP-ABC estimates can be compared to that.

I assume that you have downloaded (or cloned) the pyEPABC git-project into a folder pyEPABC_git. To run this example open an ipython shell and navigate to pyEPABC_git/examples. Then:

import pyEPABC
run testGauss

You can change parameters of EP-ABC, or the problem, e.g., the dimensionality of the data points, in testGauss.py to investigate the behaviour of EP-ABC.

This examples requires: numpy, scipy, pandas, matplotlib.

Drift-diffusion model example

This example fits a drift-diffusion model (DDM) to an example data set consisting of reaction times and choices recorded in a psychophysics experiment. The results are compared to the results produced by HDDM.

Additional to showing the working of EP-ABC on this kind of problem, this example also demonstrates how pyEPABC can be integrated with simulation functions implemented in C. The idea is that your simulation function is provided by an externally compiled shared library. This example shows how you can link this code for use in pyEPABC. Note that, if you are implementing a new model, i.e., simulation function, from scratch you may want to consider alternatives for this pure-C path. See the section on Binary Extensions in the Python Packaging User Guide for further information.

Before you can run the example you need to compile a few things. I'll only explain how to do this in Linux. First, compile the 'external' library which will provide super-fast simulations from a DDM. Do this by opening a shell and navigating to the pyEPABC_git/src folder. There run:

mkdir ../lib
gcc -shared -fPIC -O3 -lc -o ../lib/libDDMsampler.so DDMsampler.c brownian_motion_simulation.c

which will place the resulting shared library into pyEPABC_git/lib. Then, compile the Python-C-interface:

python setup.py build_ext -i -useC

which will generate another shared library testDDM_C.xxx.so where xxx will depend on your python setup. This library can be directly imported into python as a module, as done in testDDM.py. To run the example open an ipython shell, navigate to the pyEPABC_git folder and execute:

import pyEPABC
run examples/testDDM

This example requires: numpy, scipy, pandas, matplotlib, seaborn, cython, hddm.

Narrow posteriors

EP-ABC sometimes produces too narrow posteriors. See examples/narrow_posteriors.ipynb for an explanation, how to identify this issue and how to prevent it.