• pykrusch
  • Example
  • Installation Prerequisites
  • Installation
  • Usage
    • As a Function
    • From the Terminal

pykrusch

pykrusch is a package for visualizing Bayesian Generalized Linear Model structure. It is built atop pymc -- a Python package for building and fitting Bayesian statistical models.

This package was inspired by the diagrams John K. Kruschke developed to visualize Bayesian model structure in his book, Doing Bayesian Data Analysis.

Example

pykrusch is designed to produce informative, intuitive visualizations of Bayesan GLMs specified in pymc. It takes in a model specification such as the following:

import pymc as pm
import numpy as np

x = np.random.normal(loc=0, scale=1, size=10) # Data Placeholder
y = np.random.normal(loc=2*x, scale=1, size=10) # Data Placeholder


with pm.Model() as model:

	# Priors
	α = pm.Normal("α", 1, 2)
	β = pm.Normal("β", -1, 2)
	σ = pm.Exponential("σ", 3)

	# Linear Model
	μ = pm.Deterministic("μ", α + β*x)
	
	# Likelihood
	y_ = pm.Normal("y", μ, σ, observed=y)

With the pymc model specified, users may pass the pymc model to pykrusch's krusch function as the first and only argument, like so:

from pykrusch import krusch

krusch(model)

Doing so will produce a visualization of the model in the current working directory that will, by default, be titled krusch.png. Using the model code specified above, pykrusch will produce the following image:

Installation Prerequisites

Aside from an up-to-date installation of python and pip (installation instructions for which can be found here), the pykrusch package depends on graphviz, which must be installed before attempting to install pykrusch. Installation instructions for graphviz can be found at the GraphViz installation instructions page.

Installation

To install pykrusch, use pip:

pip install pkyrusch

Usage

The pykrusch package can be used as a function from within a Python (.py) / iPython (.ipynb) file, or it can be used from the terminal.

As a Function

To use pykrusch as a function, simply import the package and call the krusch() function, supplying the name of the pymc model as the first argument:

from pykrusch import krusch

krusch(model)

pykrusch will create a visualization of the Bayesian GLM it is supplied with, outputting the result as krusch.png into the working directory. By altering the outname parameter, one can change the location, name, and format of the image output.^[At this time, png is the only supported image type. Others may or may not function properly.] For example:

krusch(model, outname="img/simple_model.png")

By default, pykrusch only plots prior distributions. If one wishes to plot certain applicable posteriors alongside the priors, one can supply an arviz InferenceData object to the posterior_trace argument:

pykrusch.krusch(model, posterior_trace=trace)

The resulting is featured below, with prior distributions in blue and posterior distributions in orange:

From the Terminal

To use pykrusch from the command line, one must supply it with the name of the Python file that contains the pymc model in question.

pykrusch simple_model.py

Operating in this mode, pykrusch assumes that the name of the model in the specified file is model. If the name of the Bayesian model is not model, the model's name can be specified using the --model-name/-n argument, like so:

pykrusch simple_model.py -n simple_model

By default, pykrusch produces an output image named krusch.png. If you would like pykrusch to give the output image a different name, you can use the --outname/-o argument:^[As mentioned above, pykrusch currently only supports the use of png images as output. Most other formats will not work, and those that do may not appear exactly as the png format does.]

pykrusch simple_model.py -o model_image.png

If you would like pykrusch to plot the model's posterior distributions in addition to the priors, you have two options. The first option involves sampling from your model using pm.sample() and storing the resulting InferenceData object as a .pkl file using pickle. The resulting .pkl file can then be fed in as an argument to pykrusch's --posterior-pickle/-p argument:

pykrusch simple_model.py -p posterior_pickle.pkl

You can also instruct pykrusch to sample from the model using the default pymc sampling configuration and extract the resulting posterior distributions for plotting. This is accomplished via the --sample-posterior/-s flag:

pykrusch simple_model.py -s

Note: using the default pymc sampling behaviour may be slow and might not produce satisfactory results for all models. The sampling may contain large numbers of divergences, fail to converge, or produce nonsensical results. Only use the --sample-posterior/-s flag if you have tested your model using pm.sample() with no arguments and found the results to be to your satisfaction.