Jupyter Notebook Other
Switch branches/tags
Nothing to show
Clone or download
jonsedar Merge pull request #1 from Volodymyrk/master
rvs var was used in 01_BasicInstallTests.ipynb. conda install Pystan
Latest commit 70bd857 May 22, 2016

README.md

Jonathan Sedar Personal Project

PyMC3 Vs PyStan Comparison

Spring 2016

This set of Notebooks and scripts comprise the pymc3_vs_pystan personal project by Jonathan Sedar of Applied AI Ltd, written primarily for presentation at the PyData London 2016 Conference.

The project demonstrates hierarchical linear regression using two Bayesian inference frameworks: PyMC3 and PyStan. The project borrows heavily from code written for Applied AI Ltd and is supplied here for educational purposes only. No copyright or license is extended to users.

Copyright Applied AI Ltd 2016


Development

Git clone the repo to your workspace.

e.g. in Mac OSX terminal:

    $> git clone https://github.com/jonsedar/pymc3_vs_pystan.git
    $> cd pymc3_vs_pystan

NOTES:

  • This project uses Python 3.5, and was developed on a Macbook Pro OSX 10.10.5 using the Anaconda distro with a new virtualenv on 19 April 2016
  • The project requires PyMC3 (with an associated Theano install) and PyStan (with an associated Stan install) so is quite heavy.
  • Specific versions of key packages for clarity: pymc3-3.0, theano-0.8.1, pystan-2.9.0.0

Setup a virtual environment for Python libraries

  1. Using create a new virtualenv, installing packages from env YAML file:

     $> conda env create --file conda_env_pymc3_vs_pystan.yml
     $> source activate pymc3_vs_pystan
    
  2. install remaining packages via pip (inc pymc3 master with deps):

     $> ./pip_install.sh
    
  3. Launch Jupyter Notebook server

     $> jupyter notebook
    

Data

Local data is not stored in the repo, and should be manually copied into the subdirectory data/

See file data/README_DATA.md for more info


General Notes:

File hack_findmap.py contains a customised find_MAP() function to correct for pymc3's default behaviour of computing gradients when the chosen optimizer doesn't use them. We don't want to compute gradients in large datasets because it's quite computationally expensive.