Implementation:Pyro ppl Pyro Pyro Param

Knowledge Sources	Pyro Pyro Parameters
Domains	Probabilistic_Programming, Optimization
Last Updated	2026-02-09 00:00 GMT

Overview

Concrete tool for declaring and managing optimizable parameters in Pyro's global parameter store.

Description

pyro.param declares a named parameter that is stored in Pyro's global ParamStoreDict and optimized during SVI training. Parameters support PyTorch constraints (e.g., positivity) and are automatically transformed between constrained and unconstrained spaces. The parameter store provides persistence, serialization, and name-based lookup of all learnable parameters.

Usage

Use pyro.param to declare free parameters in custom variational guides. For most use cases, prefer AutoGuide classes (AutoNormal, AutoDelta) which manage parameters automatically. Direct param usage is needed for custom guide architectures or when fine-grained control over parameterization is required.

Code Reference

Source Location

• Repository: pyro
• File: pyro/primitives.py
• Lines: L57-91

Signature

def param(
    name: str,
    init_tensor: Union[torch.Tensor, Callable[[], torch.Tensor], None] = None,
    constraint: constraints.Constraint = constraints.real,
    event_dim: Optional[int] = None,
) -> torch.Tensor:
    """
    Saves the variable as a parameter in the param store.

    Args:
        name: name of parameter
        init_tensor: initial tensor or lazy callable
        constraint: torch constraint, defaults to constraints.real
        event_dim: optional number of rightmost dimensions unrelated to batching
    Returns:
        A constrained parameter tensor
    """

Import

import pyro
# Used as: pyro.param(name, init_tensor, ...)

I/O Contract

Inputs

Name	Type	Required	Description
name	str	Yes	Unique name for the parameter
init_tensor	Union[Tensor, Callable, None]	No	Initial value or lazy callable
constraint	constraints.Constraint	No	Constraint on parameter space (default: real)
event_dim	Optional[int]	No	Number of rightmost event dimensions

Outputs

Name	Type	Description
return	torch.Tensor	Constrained parameter tensor with gradient tracking

Usage Examples

Custom Guide Parameters

import pyro
import pyro.distributions as dist
import torch

def guide(data):
    # Learnable location and scale
    mu_loc = pyro.param("mu_loc", torch.tensor(0.0))
    mu_scale = pyro.param("mu_scale", torch.tensor(1.0),
                           constraint=dist.constraints.positive)

    pyro.sample("mu", dist.Normal(mu_loc, mu_scale))

Related Pages

Implements Principle

• Principle:Pyro_ppl_Pyro_Learnable_Parameters