stancache

author: Jacqueline Buros Novik

Overview

Filecache for stan models

Installation

You can install this package from pypi using pip:

$ pip install stancache

Or clone the repo & run setup.py:

$ git clone https://github.com/hammerlab/stancache.git
$ python setup.py install

Introduction

This is a filecache for pystan models fit to data. Each pystan model fit to data is comprised of two parts - the compiled model code & the result of MCMC sampling of that model given data. Both model compilation & model sampling can be time-consuming operations, so both are cached as separate pickled objects on the filesystem.

This separation allows one to (for example) compile a model once & execute the model several times - caching the result each time. You might be testing the model on different samples of data, or using different initializations or passing in different parameters.

Loading pickled pystan.fit objects into memory is also safer using cached_stan_fit() since this will ensure that the compiled model is first unpickled before the fit model.

Getting started

Configuratation

The configuration uses python's configparser module, allowing the user to either load a config.ini file from disk or set the configuration in code.

stancache looks for a default config file to be located in '~/.stancache.ini'. You can modify this using stancache.config.load_config('/another/config/file.ini').

Currently, the config settings include

CACHE_DIR (defaults to .cached_models)
SEED (seed value passed to pystan.stan for reproducible research)
SET_SEED (boolean, whether to set the random.seed, systemwide in addition to stan_seed)

You can use config.set_value(NAME=value) to modify a setting.

For example, you might want to set up a shared-nfs-mount containing fitted models among your collaborators:

from stancache import config
config.set_value(CACHE_DIR='/mnt/trial-analyses/cohort1/stancache')

An updated list of configuration defaults is available in defaults.py

Fitting cached models

Once you have configured your settings, you would then use stancache.cached_stan_fit to fit your model, like so:

from stancache import stancache
fit1 = stancache.cached_stan_fit(file = '/path/to/model.stan', data=dict(), chains=4, iter=100)

The options to cached_stan_fit are the same as those to pystan.stan (see pystan.stan documentation).

Also see ?stancache.cached_stan_fit for more details.

Caching other items

The caching is very sensitive to certain things which would change the returned object, such as the sort order of your data elements within the dictionary. But is not sensitive to other things (such as whether you use a file-based stan code or string-based version of same code).

In practice, we find that it can be helpful to cache data-preparation steps, especially when simulating data. There is thus as stancache.cached() wrapper function for this purpose, to cache all objects other than pystan.stan objects using the same file-cache settings.

A fairly common set-up for us is, for example, to fit a set of models in a distributed execution environment, then review the model results in a set of jupyter notebooks. In this case, in our jupyter notebook we will set a parameter of cache_only=True when loading model results into the Jupyter notebook to force a failure if the cache is not available.

Contributing

TBD

Examples

For example (borrowing from pystan's docs):

import stancache

schools_code = """
data {
    int<lower=0> J; // number of schools
    real y[J]; // estimated treatment effects
    real<lower=0> sigma[J]; // s.e. of effect estimates
}
parameters {
    real mu;
    real<lower=0> tau;
    real eta[J];
}
transformed parameters {
    real theta[J];
    for (j in 1:J)
    theta[j] <- mu + tau * eta[j];
}
model {
    eta ~ normal(0, 1);
    y ~ normal(theta, sigma);
}
"""

schools_dat = {'J': 8,
               'y': [28,  8, -3,  7, -1,  1, 18, 12],
               'sigma': [15, 10, 16, 11,  9, 11, 10, 18]}

# fit model to data
fit = stancache.cached_stan_fit(model_code=schools_code, data=schools_dat,
                                iter=1000, chains=4)

# load fit model from cache
fit2 = stancache.cached_stan_fit(model_code=schools_code, data=schools_dat,
                                 iter=1000, chains=4)

In addition, there are a number of publicly-accessible ipynbs using stancache.

These include:

If you know of other examples, please let us know and we will add them to this list.