mu-PPL is a micro probabilistic programming language (PPL) on top of Python (mu is for the greek letter µ). Following recent PPLs (WebPPL, Pyro, Gen.jl) mu-PPL provides the following probabilistic operators:
sampleto draw a sample from a distributionassumeto condition the model on a boolean condition (hard conditionning)factor/observeto update the score of the current execution (soft conditionning)
mu-PPL can be installed locally with:
$ pip install git+https://github.com/gbdrt/mu-pplHere is the mu-PPL program that computes the bias of coin from a series of observations.
from typing import List
from mu_ppl import *
def coin(obs: List[int]) -> float:
p = sample(Uniform(0, 1), name="p")
for o in obs:
observe(Bernoulli(p), o)
return p
with ImportanceSampling(num_particles=10000):
dist: Categorical[float] = infer(coin, [0, 0, 0, 0, 0, 0, 0, 0, 1, 1]) # type: ignore
print(dist.stats())You should get the following output:
100%|███████████████████████████████████| 10000/10000 [00:02<00:00, 3785.50it/s]
(0.2511526566307631, 0.11951005324220002)More examples are available in the examples directory.
The following introductory notebooks can be run locally or in Google Colab.
