Bayesian Non-parametrics Playground

A little playground for defining and running Bayesian nonparametric models, in the style of Roy et al. (2008) and LazyPPL.

Note that this is currently not a library for fitting these models, just for running them to generate synthetic data. It is intended for pedagogical purposes & quick prototyping.

Getting started

You can install the package and run a test script using poetry.

poetry install
poetry run python examples/examples.py

The package provides several basic components from which more complex models can be built:

1. InfiniteArray(f): given a stochastic function f from indices to values, sample an infinite array of values by calling f(i) on every index. (In practice, this works via stochastic memoization: if a = InifiniteArray(f), then the first time a[i] is accessed for a new index i, f(i) is called; subsequent lookups of a[i] then always return the same value.)
2. GEM(alpha): generates an infinite vector of probabilities from a GEM (Griffiths, Engen, McCloskey) distribution.
3. DP(alpha, H): generates a random probability measure with base measure H (a stochastic function).

It also provides examples of using these components to define various models, including DPMMs, IRMs, CrossCat models, and hierarchical IRMs.