Auxiliary pseudo-marginal MCMC python implementations
Python implementations of MCMC samplers in the auxiliary pseudo-marginal MCMC
framework as described in the paper
Pseudo-Marginal Slice Sampling and
associated code for running Gaussian process classification model parameter
inference experiments.
A simpler single module Python implementation written by Iain Murray is also available here - this is probably the simplest option for applying the method to your own problem.
The code has only been tested in
Python 2.7 and there are no
guarantees it will work at all in other Python versions.
Minimal requirements for using the provided package are:
numpy (1.9.2)scipy (0.16.0) (only required for gpdemo package)matplotlib (1.4.3) (only required for gpdemo package)The versions specified are those the code was developed and tested on -
different versions may work as well.
To build the Cython modules yourself rather than using the
pre-built C-code you will also need Cython (0.22).
For viewing the IPython notebooks for running the
experiments and analysing the results you will also need to have a workingIPython (3.1.0) install and all the dependencies for the
IPython notebook server. For example run
pip install ipython[notebook]
For the results analysis you will need to have a system R installation and
also have rpy2 (2.7.0) a python — R interface installed.
Run python setup.py install from main package directory to install the
package into the currently active python environment. This will also
build the Cython modules in the package from the provided C-source.
If you have Cython installed you can also specify for the Cython code to be
built directly from the Cython source by instead running
python setup.py install -use-cython
For other install options refer run python setup.py --help
The code is organised in to three main sub-directories:
auxpm
The Python package containing the modules implementing the different
auxiliary pseudo-marginal samplers variants (auxpm.samplers) and MCMC
update steps (auxpm.mcmc_updates).
gpdemo
The Python package containing the modules implementing the functions
specific to the Gaussian process classification parameter inference
experiements.
experiment_notebooks
A series of IPython notebooks using the above two packages to run
Gaussian process classification parameter inference experiments for
different sampling methods and analyse results.
To run the experiment notebooks a local copy of any or all of the UCI
classification datasets used in the experiments in the paper
Filippone, Maurizio, and Mark Girolami.
‘Pseudo-marginal Bayesian inference for Gaussian processes.’
Pattern Analysis and Machine Intelligence,
IEEE Transactions on 36.11 (2014): 2214-2226.
will be required. These can be downloaded in the requisite space-delimited text
file format as part of the code associated with that paper at
http://www.dcs.gla.ac.uk/~maurizio/Code/code_pseudo_marg.tar.gz
The data files are in the section4.4/DATA/clean sub-directory of the archive.
Each dataset has a text file with suffix _X.txt containing the input features
and _y.txt suffice containing the targets.
These datasets were originally taken from the UCI Machine Learning Repository
Lichman, M. (2013).
UCI Machine Learning Repository http://archive.ics.uci.edu/ml.
Irvine, CA: University of California,
School of Information and Computer Science.
The relevant text data files should be placed under a uci sub-directory which
itself is placed in a directory readable by the current user and the path to
which is specified in a environment variable DATA_DIR defined for the current
user. So for example if the Pima Indians dataset files are taken from the
archive linked to above, the inputs pima_X.txt and outputs pima_y.txt files
should exist respectively at
$DATA_DIR/uci/pima_X.txt$DATA_DIR/uci/pima_y.txt
assuming Unix type directory separators and environment variable syntax.
When running the experiment notebooks it is expected that a further EXP_DIR
environment variable will be defined for the current user which specifies a
path writeable by the current user to output experiment results to, with
results being placed under a sub-directory apm_mcmc which should be created
before running any of the notebooks.