Implementation:Pyro ppl Pyro Autoname Scoping

Property	Value
Implementation Type	Pattern Doc
Source File	examples/contrib/autoname/scoping_mixture.py
Module	pyro.contrib.autoname
Pyro Features	pyro.contrib.autoname.scope, config_enumerate, TraceEnum_ELBO, pyro.plate, discrete enumeration
Pattern	Gaussian Mixture Model using scope decorator for site name prefixing

Overview

This file demonstrates using the @scope decorator from pyro.contrib.autoname to build a Gaussian Mixture Model with modular local model/guide functions. The @scope(prefix="local") decorator automatically prefixes all pyro.sample and pyro.param site names within the decorated function, enabling clean separation of global and local model components without manual name management.

Unlike the named.Object approach in the companion mixture example, the scoping approach works with discrete enumeration via config_enumerate and TraceEnum_ELBO, allowing exact marginalization of the discrete mixture assignments.

Code Reference

def model(K, data):
    weights = pyro.param("weights", torch.ones(K) / K, constraint=constraints.simplex)
    locs = pyro.param("locs", 10 * torch.randn(K))
    scale = pyro.param("scale", torch.tensor(0.5), constraint=constraints.positive)

    with pyro.plate("data"):
        return local_model(weights, locs, scale, data)

@scope(prefix="local")
def local_model(weights, locs, scale, data):
    assignment = pyro.sample("assignment",
        dist.Categorical(weights).expand_by([len(data)]))
    return pyro.sample("obs", dist.Normal(locs[assignment], scale), obs=data)

def guide(K, data):
    assignment_probs = pyro.param("assignment_probs",
        torch.ones(len(data), K) / K, constraint=constraints.unit_interval)
    with pyro.plate("data"):
        return local_guide(assignment_probs)

@scope(prefix="local")
def local_guide(probs):
    return pyro.sample("assignment", dist.Categorical(probs))

I/O Contract

Parameter	Type	Description
K	int	Number of mixture components
data	torch.Tensor	1D tensor of observed data points
-n / --num-epochs	int	Number of training epochs (default: 200)

Named sites (after scoping):

• local/assignment: Categorical component assignment (auto-prefixed)
• local/obs: Normal observation (auto-prefixed)
• weights, locs, scale: Global parameters
• assignment_probs: Guide parameters

Usage Examples

import pyro
from pyro.infer import SVI, TraceEnum_ELBO, config_enumerate

pyro.set_rng_seed(0)
pyro.clear_param_store()
K = 2
data = torch.tensor([0.0, 1.0, 2.0, 20.0, 30.0, 40.0])

optim = pyro.optim.Adam({"lr": 0.1})
inference = SVI(model, config_enumerate(guide), optim,
                loss=TraceEnum_ELBO(max_plate_nesting=1))

for step in range(200):
    loss = inference.step(K, data)

Related Pages

• Implementation:Pyro_ppl_Pyro_Autoname_Mixture - Mixture model using named.Object approach
• Implementation:Pyro_ppl_Pyro_Autoname_TreeData - Hierarchical model using named objects