Stan Model Server

NOTE: Stan Model Server is no longer under active development. For a more efficient interface to a Stan model's methods from multiple languages, check out bridgestan.

A lightweight server interface to Stan model methods.

Installation Instructions

For installation instructions, see

Installing Stan Model Server Install Documentation

Read-Eval-Print Loop

For full documentation on the REPL, see:

Stan Model Server REPL Documentation

Hello, Shell!

The package is distributed with a a Stan program in stan/bernoulli/bernoulli.hpp and matching data in stan/bernoulli/bernoulli.data.json.

To compile the server executable,

$ cd stan-model-server
$ make CMDSTAN=/Users/carp/github/stan-dev/cmdstan/ stan/bernoulli/bernoulli

Then we run from the command line with data and an RNG seed.

$ stan/bernoulli/bernoulli -s 1234 -d stan/bernoulli/bernoulli.data.json

The lines marked with < indicate input from the user and the unmarked lines are responses from the server.

< name
bernoulli_model

< param_num 1 1
3

< param_unc_num
1

< param_names 1 1
theta,logit_theta,y_sim

< param_unc_names
theta

< param_constrain 1 1 -2.3
0.0911229610148561,-2.3,0

< param_unconstrain {"theta":0.09112}
-2.30003575312272

< log_density 1 1 1 1 -1.5
-6.91695933579303,0.810893714323724,-1.78975742485563

Motivation and Inspiration

We want to be able to access Stan model methods from within R or Python in order to do algorithm development. The first system that did this is HTTPStan, an official Stan project:

GitHub stan-dev: httpstan

The direct inspiration for this project came from the simple I/O structure of ReddingStan:

Dan Muck and Daniel Lee. 2022. Smuggling log probability and gradients out of Stan programs — ReddingStan. The Stan Blog.
Github: dmuck/redding-stan

Python Client

The Python client is feature complete. It can be invoked this way after the Bernoulli example is compiled.

> import StanModelClient as smc
> sc = smc.StanClient("./stan/bernoulli/bernoulli",
                      data = "stan/bernoulli/bernoulli.data.json",
                      seed = 1234)

Here are some example calls.

> sc.name()

> sc.param_names()

> sc.log_density_gradient([-2.3])

> sc.param_constrain([-2.3])

Documentation is available as docstrings in the source code.

Python-based Samplers

The plan is to build samplers out in Python using this interface. The samplers are even more a work-in-progress than the client interface. For now, there is a worked example of Metropolis with a Stan model.

R Client

The R Client is feature complete. It can be invoked this way after the Bernoulli example is compiled.

# Load source file into environment
source('StanModelClient.R')

# Create the Stan Client 
sc <- create_stan_client(exe_file = "./stan/bernoulli/bernoulli", 
                         data_file = "stan/bernoulli/bernoulli.data.json", 
                         seed = 1234)

Here are some example calls.

sc$name()

sc$param_names()

sc$log_density_gradient(-2.3)

sc$param_constrain(-2.3)

For more information on the R Client, see the source code and the documentation for a complete list of commands and their usage.

R-based Samplers

In time, the samplers available in Python will also be available through the R interface. For now, however, see the worked example of gradient descent on the Bernoulli Stan model.

License

Code released under the BSD-3 license.
Documentation released under the CC BY 4.0 license.

bob-carpenter / stan-model-server