How can Shogun interface STAN?

[karlnapf]

This task is on the GSoC 2014 project http://www.shogun-toolbox.org/page/Events/gsoc2014_ideas#mcmc Stan webpage: http://mc-stan.org/

While the project is about writing a modular class representation of MCMC algorithms in Shogun, we would like to avoid actually implementing the samplers on our own: This is cumbersome work and very challenging. Rather we would like to interface the STAN library for MCMC, which contains very mature implementations of most interesting samplers.

This entrance task is about finding out how this interfacing would work.

• How can we interface STAN from c++? Are internal data representations compatible?
• How could one call STAN samplers from Shogun. How to call/run samplers on fixed targets in STAN?
• How to represent target densities implemented as Shogun functions in STAN?
• Could you produce a mini-running example? Optimally, one would want to pass a Shogun object with a log_pdf method to a handler that then sets up STAN and for example runs standard Metropolis-Hastings.

The open-ended task is a lot about exploring some of the internals of both Shogun and STAN. Some basic understanding of STAN would be very helpful. If you can up with an elegant idea to do this (and maybe an example) then this would be a big big plus for your application on this project.

The GSoC project idea description might contain some useful references.

[karlnapf]

See some simple example how to interface stan's distribution implementation from c++ This is a first step. https://gist.github.com/karlnapf/9354664

[karlnapf]

In addition, a very important aspect (and subtask) is to integrate stan's source into Shogun.

• have a cmake option to bundle stan or to provide an installation to link against
• stan comes with its own eigen3, we however want to use Shogun's
• stan also depends on boost, so in case stan is bundled, we need to add a boost dependency
• add linker and compiler flags if it is used
• include guards
• etc, see for example the way eigen3 is bundled.

@sonney2k and @vigsterkr might have comments here

[vigsterkr]

@karlnapf have you checked the code of stan? it's a horror. i mean distribution wise. the repository contains full src of boost, eigen (and something else i cannot remember now). we _do not_ want to bundle this into shogun. it would like bloat shogun's footprint by another 500megs or so....

[karlnapf]

@vigsterkr mmmmh ok. What else could we do there? I just want to be able to access some of stan's code for automatic differentiation, distributions and mcmc. But the result can also be negative, just investigating.....

[souravsingh]

@karlnapf Would the project still be relevant for shogun right now?

[karlnapf]

Definitely, but this is a big undertaking. Better start smaller