vishalbelsare / pm-pyro Goto Github PK

PyMC3-like abstractions for pyro's stochastic function. Define a model as a stochastic function in pyro. Use pm_like wrapper to create a PyMC3-esque Model. Random variables are exposed to user as attributes of Model. pm-pyro provides abstractions for sampling-based inference methods (NUTS - The No-U-Turn Sampler, HMC - Hamiltonion Monte Carlo), as well as Variational Inference (SVI with autoguides), trace plots, posterior plot and posterior predictive plots.

Install from pypi

pip install pm-pyro

Developer setup

# install requirements
pip install -r requirements-dev.txt
# run tests
python -m pytest pmpyro/tests.py

Borrowed the example from a PyMC3 tutorial. The outcome variable Y is dependent on 2 features X_1 and X_2. The notebook for this example is available here

We design a simple Bayesian Linear Regression model.

The model specification is implemented as a stochastic function.

import pyro.distributions as dist
import pyro
import torch

def pyro_model(x1, x2, y):
    alpha = pyro.sample('alpha', dist.Normal(0, 10))
    beta = pyro.sample('beta',pdist.Normal(torch.zeros(2,), torch.ones(2,) * 10.))
    sigma = pyro.sample('sigma', dist.HalfNormal(1.))

    # Expected value of outcome
    mu = alpha + beta[0] * x1 + beta[1] * x2

    # Likelihood (sampling distribution) of observations
    return pyro.sample('y_obs', dist.Normal(mu, sigma), obs=y)

The pm_like wrapper creates a PyMC3-esque Model. We can use the context manager syntax for running inference. pm.sample samples from the model using the NUTS sampler. The trace is a python dictionary which contains the samples.

from pmpyro import pm_like
import pmpyro as pm

with pm_like(pyro_model, X1, X2, Y) as model:
    trace = pm.sample(1000)

sample: 100%|██████████| 1300/1300 [00:16, 80.42it/s, step size=7.49e-01, acc. prob=0.911] 

We can visualize the samples using traceplot. Select random variables by passing them as a list via var_names = [ 'alpha' ... ] argument.

pm.traceplot(trace)

Visualize posterior of random variables using plot_posterior.

pm.plot_posterior(trace, var_names=['beta'])

We can sample from the posterior by running plot_posterior_predictive or sample_posterior_predictive with the same function signatures as the stochastic function def pyro_model(x1, x2, y), replacing observed variable Y with None.

ppc = pm.plot_posterior_predictive(X1, X2, None,
                          trace=trace, model=model, samples=60,
                          alpha=0.08, obs={'y_obs' : Y})

The summary of random variables is available as a pandas array.

pm.summary()

This project is licensed under the GPL v3 License - see the LICENSE.md file for details