Implementation:Pyro ppl Pyro MiniPyro Example

Property	Value
Implementation Type	Pattern Doc
Source File	examples/minipyro.py
Module	pyro.contrib.minipyro
Pyro Features	pyro.generic interface, pyro_backend, SVI, Trace_ELBO, JitTrace_ELBO, backend switching
Pattern	Demonstrates MiniPyro -- a minimal PPL for didactic purposes

Overview

This file demonstrates pyro.contrib.minipyro, a minimal implementation of the Pyro Probabilistic Programming Language created for educational purposes. The example shows that MiniPyro's API is compatible with the full Pyro API, so the same model, guide, and training code works with both backends.

The example uses the pyro.generic interface, which provides a dynamic backend switching mechanism via pyro_backend(). This means you can write probabilistic programs that work with:

• "minipyro": The minimal backend (for learning/teaching)
• "pyro": The full Pyro backend (for production)

The model is a simple conjugate Gaussian: a latent location loc ~ Normal(0, 1) observed through Normally distributed data obs ~ Normal(loc, 1). The guide is a diagonal Normal with learnable mean and scale.

Code Reference

from pyro.generic import distributions as dist
from pyro.generic import infer, ops, optim, pyro, pyro_backend

def main(args):
    def model(data):
        loc = pyro.sample("loc", dist.Normal(0.0, 1.0))
        with pyro.plate("data", len(data), dim=-1):
            pyro.sample("obs", dist.Normal(loc, 1.0), obs=data)

    def guide(data):
        guide_loc = pyro.param("guide_loc", torch.tensor(0.0))
        guide_scale = ops.exp(pyro.param("guide_scale_log", torch.tensor(0.0)))
        pyro.sample("loc", dist.Normal(guide_loc, guide_scale))

    data = torch.randn(100) + 3.0

    with pyro_backend(args.backend):
        Elbo = infer.JitTrace_ELBO if args.jit else infer.Trace_ELBO
        elbo = Elbo()
        adam = optim.Adam({"lr": args.learning_rate})
        svi = infer.SVI(model, guide, adam, elbo)
        for step in range(args.num_steps):
            loss = svi.step(data)

I/O Contract

Parameter	Type	Description
-b / --backend	str	Backend to use: "minipyro" or "pyro" (default: "minipyro")
-n / --num-steps	int	SVI optimization steps (default: 1001)
-lr / --learning-rate	float	Learning rate (default: 0.02)
--jit	flag	Use JIT-compiled ELBO

Output:

• Loss at every 100 steps
• Final learned variational parameters (guide_loc should be ~3.0)
• Assertion that guide_loc is within 0.1 of 3.0

Usage Examples

# Run with MiniPyro backend (default)
# python minipyro.py -n 1001 -lr 0.02

# Run with full Pyro backend
# python minipyro.py -b pyro -n 1001

# Run with JIT
# python minipyro.py --jit -b pyro

Related Pages

• Pyro_ppl_Pyro_InclinedPlane - Another introductory inference example