Skip to content

WardBrian/cookiecutter-cmdstanpy-wrapper

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

cookiecutter-cmdstanpy-wrapper

This project a cookiecutter template designed for package developers who wish to wrap a Stan model in a Python package. It provides the ability to pre-compile the model as part of a Python wheel, removing the need for end users to have a C++ toolchain installed to install or use the package.

This uses cmdstanpy as the underlying interface to the Stan model.

If you're looking instead for a template for doing applied modeling work in CmdStanPy, check out cookiecutter-cmdstanpy-analysis.

See an example based on the output of the template here

Usage

cookiecutter gh:WardBrian/cookiecutter-cmdstanpy-wrapper

This will ask you some basic prompts and then generate a folder with skeleton of a Python package. This package will include the setup.py required for building the models as part of the wheel, and a Github Actions script to do so on MacOS, Linux, and Windows using cibuildwheel.

How it works

The process of shipping a pre-compiled Stan model is a bit more complicated than it needs to be due to Stan's reliance on Intel's TBB library.

To make a Stan model work in the most generality (e.g., on a machine it was not compiled on), it needs to also have a compiled dynamic shared object for TBB. This is done by (optionally) repackaging part of the CmdStan distribution alongside the built model. This allows TBB to be detected on the new system. The built-in relocation tools in cibuildwheel ensure this DSO is relocatable and that the compiled Stan models work.

The above behavior is only enabled when an environment variable called PKG_NAME_REDISTRIBUTE_CMDSTAN is set. Source installations assume that the end user has a working CmdStan distribution already.

License

This cookiecutter and the building code is licensed under the MIT license.

This project is modeled on work I assisted with on Facebook's prophet package, and the yet-unreleased scitkit-stan package.