pymc3-quap

The quadratic approximation is a very fast method to approximate the posterior with a multivariate normal.

NOTE: The quadratic approximation only works well if the posterior is uni-modal and roughly symmetrical.

Example

import numpy as np
import pymc3 as pm
import arviz as az
from quap import quap
y = np.array([2642, 3503, 4358]*10)

# Normal with unknown mean and log-variance, with uniform priors 
with pm.Model() as m: 
  logsigma = pm.Uniform("logsigma", -100, 100)
  mu = pm.Uniform("mu", -10000, 10000) 
  yobs = pm.Normal("y", mu=mu, sigma=pm.math.exp(logsigma), observed=y)
  idata, posterior = quap([mu, logsigma])

az.plot_posterior(idata)

idata is an arviz.InferenceData with samples from the approximate posterior for compatibility with the Arviz ecosystem.

posterior is the exact approximate posterior scipy.stats.multivariate_normal

True posterior and quadratic approximation for the example above.

Install

pip install pymc3-quap