pycmdstan

Python interface to CmdStan.

Install

Pycmdstan is a pure-Python package which can be installed from the Git repo

pip install -e git+https://gitlab.thevirtualbrain.org/tvb/pycmdstan

or (eventually) from PyPI

pip install pycmdstan

Usage

import os
os.environ['CMDSTAN'] = '~/src/cmdstan-2.17.1'
from pycmdstan import Model, Run

model = Model('''
data { vector[20] x; real mu; }
parameters { real sig; }
model { x ~ normal(mu, sig); }
generate quantities {
    vector[20] log_lik;
    for (i in 1:20) log_lik[i] = normal_lpdf(x[i] | mu, sig);
}
''')

base_data = {'x': np.random.randn(20) + 5.0}

loo = []
for mu in np.r_[1.0, 3.0, 5.0, 7.0, 9.0]:
    data = {'mu': mu}
    data.update(base_data)
    run = Run(
        model,
        'sample',
        data,
        method_args={
            'num_warmup': 200,
            'num_samples': 200
        })
    loo.append(run['loo'])
loo = np.array(loo)

a notable goal is to be able to inspect warmup while it's running.

Roadmap (wishlist)

• preprocess & compile Stan model with support for #includes
• run model w/ multiple chains
• invoke summary or other executables
• read/write Rdump format (modulo a bug somewhere)
• read sample & summary CSV format
• model comparison, at least PSIS LOO
• common visualizations (trace, pairs, rhat hist)
• coordinators opt_to_init, sim_fit which can sequence model invocations for common cases
• typical makefile suspects invoke_stanc, compile_model, etc
• bind standalone funcs via ctypes
• bind density functions + jupyter widgets to tune dist params interactively
• auto get cmdstan sources & build

Dev

Use YAPF to format the code. The Dockerfile can ease local development,

docker build -t pycmdstan .
docker run --rm -it -v `pwd`:/opt/pycmdstan pycmdstan python -m unittest discover

Acknowledgements

• PSIS code is by Aki Vehtari & Tuomas Sivula (BSD licensed, repo here)