golem

Golem is a minimalistic framework for performing Bayesian analysis. It was heavily inspired by Richard McElreath's Statistical Rethinking course, which is a lovely journey into the scientific thinking. It's built on top of PyTorch, and more specifically, it uses the torch.distributions module and the available optimization routines.

My main goal is to have a simple interface to sketch ideas and test them quickly. Caution is advise as the tool should be mainly used didactically to get a better grasp of the Bayesian paradigm.

Installation

Install PyTorch and then run:

pip install -e .

Usage

Let's assume we want to model the data collected by Howell between August 1967 and May 1969 in Botswana of the !Kung San tribe (read more here).

import pandas as pd
from golem import Prior, LinearCombination, Model, Golem

# Load the data
data = pd.read_excel('AdultHtWtOne_each_17985.xls')[['htcms', 'wtkgs']].dropna()
H = torch.tensor(data['htcms'].values)
W = torch.tensor(data['wtkgs'].values)

# Define the priors
intercept_prior = Prior('intercept', torch.distributions.Normal, loc=178, scale=20)
slope_prior     = Prior('slope', torch.distributions.LogNormal, loc=0, scale=1)
sigma_prior     = Prior('sigma', torch.distributions.Uniform, low=0, high=50)
mu_prior        = LinearCombination(H, intercept_prior, slope_prior)

# Define the Model
model = Model(W, torch.distributions.Normal, loc=mu_prior, scale=sigma_prior)

# Conjure the Golem
golem = Golem([model])

# Before sampling, we need to estimate the MAP to initialize the sampler
golem.maximum_a_posteriori()

# Sample from the posterior
samples = golem.sample(n_samples=10000, burn_in=5000)

Now that we have the distribution associated with the parameters of the model, the sky is the limit.