Implementation:Pyro ppl Pyro Distributions Init

Attribute	Value
Sources	pyro/distributions/__init__.py
Domains	Probabilistic Programming, Distribution Registry, Module Initialization
Last Updated	2026-02-09

Overview

Description

The pyro/distributions/__init__.py module serves as the central registry and public API entry point for all probability distributions available in Pyro. It aggregates distributions from three sources:

1. PyTorch upstream distributions -- Re-exported from pyro.distributions.torch, which wraps all of torch.distributions with Pyro-specific enhancements. These include standard distributions like Normal, Bernoulli, Beta, Categorical, Dirichlet, and many others.
2. Pyro-native distributions -- Custom distributions implemented specifically for Pyro, covering specialized domains like directional statistics, hidden Markov models, coalescent processes, zero-inflated models, and heavy-tailed distributions.
3. Infrastructure classes -- Foundational classes such as TorchDistribution, Distribution, TransformModule, and utility functions for validation control.

The module applies a patch to PyTorch distributions (pyro.distributions.torch_patch) on import, then performs a wildcard import from the torch wrapper module, followed by explicit imports of all Pyro-native distributions. The final __all__ list is constructed by combining the Pyro-specific names with the torch distribution names, sorting and deduplicating the result.

Additionally, the module exports three submodules: constraints, kl (Kullback-Leibler divergences), and transforms.

Usage

Users typically import distributions from this top-level module: from pyro.distributions import Normal, Stable, SineSkewed. This single import path provides access to all available distributions regardless of whether they originate from PyTorch or are Pyro-native implementations.

Code Reference

Source Location

Property	Value
File	pyro/distributions/__init__.py
Module	pyro.distributions
Repository	pyro-ppl/pyro

Signature

# Core initialization pattern
import pyro.distributions.torch_patch      # Apply PyTorch patches
from pyro.distributions.torch import *      # Re-export all PyTorch distributions

# Explicit PyTorch distribution imports (for mypy)
from pyro.distributions.torch import (
    Bernoulli, Beta, Binomial, Categorical, Cauchy, Chi2,
    ContinuousBernoulli, Dirichlet, Exponential, ...
)

# Pyro-native distribution imports
from pyro.distributions.affine_beta import AffineBeta
from pyro.distributions.stable import Stable, StableWithLogProb
from pyro.distributions.unit import Unit
# ... (many more)

# Submodule exports
from . import constraints, kl, transforms

# Combined __all__ list
__all__ = [...]  # 80+ distribution names, sorted and deduplicated

Import

# Standard user-facing import
import pyro.distributions as dist

# Or import specific distributions
from pyro.distributions import Normal, Stable, SineSkewed, Unit

I/O Contract

Exported PyTorch Distributions

Distribution	Description
Bernoulli	Bernoulli distribution for binary outcomes
Beta	Beta distribution on [0, 1]
Binomial	Binomial distribution for count data
Categorical	Categorical distribution over discrete categories
Cauchy	Cauchy (Lorentz) distribution
Chi2	Chi-squared distribution
ContinuousBernoulli	Continuous relaxation of Bernoulli
Dirichlet	Dirichlet distribution over simplexes
Exponential	Exponential distribution
FisherSnedecor	F-distribution
Gamma	Gamma distribution
Geometric	Geometric distribution
Gumbel	Gumbel (extreme value type I) distribution
HalfCauchy	Half-Cauchy distribution
HalfNormal	Half-Normal distribution
Independent	Wraps a distribution to reinterpret batch dims as event dims
Kumaraswamy	Kumaraswamy distribution on [0, 1]
Laplace	Laplace distribution
LKJCholesky	LKJ distribution on Cholesky factors of correlation matrices
LogisticNormal	Logistic-Normal distribution
LogNormal	Log-Normal distribution
LowRankMultivariateNormal	Multivariate normal with low-rank plus diagonal covariance
MixtureSameFamily	Mixture distribution with same family components
Multinomial	Multinomial distribution
MultivariateNormal	Multivariate normal distribution
NegativeBinomial	Negative binomial distribution
Normal	Normal (Gaussian) distribution
OneHotCategorical	One-hot categorical distribution
OneHotCategoricalStraightThrough	One-hot categorical with straight-through gradient
Pareto	Pareto distribution
Poisson	Poisson distribution
RelaxedBernoulli	Relaxed (Gumbel-Softmax) Bernoulli distribution
RelaxedOneHotCategorical	Relaxed (Gumbel-Softmax) categorical distribution
StudentT	Student's t-distribution
TransformedDistribution	Distribution formed by applying transforms
Uniform	Uniform distribution
VonMises	Von Mises distribution on the circle
Weibull	Weibull distribution
Wishart	Wishart distribution on positive-definite matrices

Exported Pyro-Native Distributions

Distribution	Source Module	Description
AffineBeta	affine_beta	Beta distribution with affine transformation
AsymmetricLaplace	asymmetriclaplace	Asymmetric Laplace distribution
SoftAsymmetricLaplace	asymmetriclaplace	Smooth asymmetric Laplace distribution
AVFMultivariateNormal	avf_mvn	Auxiliary variational function multivariate normal
CoalescentRateLikelihood	coalescent	Coalescent rate likelihood
CoalescentTimes	coalescent	Distribution over coalescent times
CoalescentTimesWithRate	coalescent	Coalescent times with rate parameter
ConditionalDistribution	conditional	Base class for conditional distributions
ConditionalTransform	conditional	Conditional transform
ConditionalTransformedDistribution	conditional	Conditional transformed distribution
ConditionalTransformModule	conditional	Learnable conditional transform
BetaBinomial	conjugate	Beta-Binomial conjugate distribution
DirichletMultinomial	conjugate	Dirichlet-Multinomial conjugate distribution
GammaPoisson	conjugate	Gamma-Poisson (Negative Binomial) conjugate distribution
Delta	delta	Delta (point mass) distribution
MixtureOfDiagNormals	diag_normal_mixture	Mixture of diagonal normals
MixtureOfDiagNormalsSharedCovariance	diag_normal_mixture_shared_cov	Mixture of diagonal normals with shared covariance
Distribution	distribution	Pyro base distribution class
Empirical	empirical	Empirical distribution from samples
ExtendedBetaBinomial	extended	Extended Beta-Binomial distribution
ExtendedBinomial	extended	Extended Binomial distribution
FoldedDistribution	folded	Folded (absolute value) distribution
GaussianScaleMixture	gaussian_scale_mixture	Gaussian scale mixture distribution
GroupedNormalNormal	grouped_normal_normal	Grouped Normal-Normal model
DiscreteHMM	hmm	Discrete Hidden Markov Model
GammaGaussianHMM	hmm	Gamma-Gaussian HMM
GaussianHMM	hmm	Gaussian HMM
GaussianMRF	hmm	Gaussian Markov Random Field
IndependentHMM	hmm	Independent HMM
LinearHMM	hmm	Linear HMM
ImproperUniform	improper_uniform	Improper uniform distribution
InverseGamma	inverse_gamma	Inverse gamma distribution (Pyro fallback if not in PyTorch)
LKJ	lkj	LKJ distribution on correlation matrices
LKJCorrCholesky	lkj	LKJ on Cholesky factors of correlation matrices
LogNormalNegativeBinomial	log_normal_negative_binomial	Log-Normal Negative Binomial
Logistic	logistic	Logistic distribution
SkewLogistic	logistic	Skew logistic distribution
MaskedMixture	mixture	Masked mixture of two distributions
MultivariateStudentT	multivariate_studentt	Multivariate Student's t-distribution
NanMaskedMultivariateNormal	nanmasked	NaN-masked multivariate normal
NanMaskedNormal	nanmasked	NaN-masked normal
OMTMultivariateNormal	omt_mvn	Orthogonal Monte Carlo multivariate normal
OneOneMatching	one_one_matching	One-to-one matching distribution
OneTwoMatching	one_two_matching	One-to-two matching distribution
OrderedLogistic	ordered_logistic	Ordered logistic distribution
TruncatedPolyaGamma	polya_gamma	Truncated Polya-Gamma distribution
ProjectedNormal	projected_normal	Projected normal distribution on the sphere
Rejector	rejector	Rejection sampling distribution
RelaxedBernoulliStraightThrough	relaxed_straight_through	Relaxed Bernoulli with straight-through gradient
RelaxedOneHotCategoricalStraightThrough	relaxed_straight_through	Relaxed one-hot categorical with straight-through gradient
SineBivariateVonMises	sine_bivariate_von_mises	Sine bivariate von Mises distribution
SineSkewed	sine_skewed	Sine-skewed torus distribution
SoftLaplace	softlaplace	Smooth Laplace-like distribution
SpanningTree	spanning_tree	Distribution over spanning trees
Stable	stable	Levy alpha-stable distribution
StableWithLogProb	stable	Stable distribution without automatic reparameterization
VonMises3D	von_mises_3d	Von Mises-Fisher distribution in 3D
ZeroInflatedDistribution	zero_inflated	Zero-inflated wrapper distribution
ZeroInflatedNegativeBinomial	zero_inflated	Zero-inflated negative binomial
ZeroInflatedPoisson	zero_inflated	Zero-inflated Poisson

Exported Infrastructure Classes

Name	Source Module	Description
TorchDistribution	torch_distribution	Base class for Pyro-compatible distributions wrapping PyTorch
ExpandedDistribution	torch_distribution	Distribution with expanded batch dimensions
MaskedDistribution	torch_distribution	Distribution with masked log_prob
TransformModule	torch_transform	Base class for learnable transforms (nn.Module + Transform)
ComposeTransformModule	torch_transform	Composition of learnable transforms
Unit	unit	Trivial distribution for pyro.factor statements

Exported Utilities

Name	Description
enable_validation	Enable or disable distribution argument validation globally
is_validation_enabled	Check whether validation is currently enabled
validation_enabled	Context manager for temporarily enabling/disabling validation

Exported Submodules

Submodule	Description
constraints	Pyro-specific distribution constraints (extends PyTorch constraints)
kl	Kullback-Leibler divergence implementations for Pyro distributions
transforms	Distribution transforms including normalizing flow components

Usage Examples

Importing Distributions

# Recommended: import the module as dist
import pyro.distributions as dist

# Use PyTorch distributions seamlessly
normal = dist.Normal(0.0, 1.0)
beta = dist.Beta(2.0, 5.0)

# Use Pyro-native distributions the same way
stable = dist.Stable(stability=1.5, skew=0.3)
soft_laplace = dist.SoftLaplace(loc=0.0, scale=1.0)

Importing Specific Distributions

from pyro.distributions import (
    Normal,
    Stable,
    SineSkewed,
    SineBivariateVonMises,
    Unit,
    TransformModule,
)

Using Constraints and Transforms

import pyro.distributions as dist

# Access the constraints submodule
positive_constraint = dist.constraints.positive
simplex_constraint = dist.constraints.simplex

# Access the transforms submodule
exp_transform = dist.transforms.ExpTransform()

Validation Control

import pyro.distributions as dist

# Enable validation globally
dist.enable_validation(True)
print(dist.is_validation_enabled())  # True

# Use context manager for temporary validation
with dist.validation_enabled(False):
    # Validation disabled in this block
    d = dist.Normal(0.0, -1.0)  # Would normally raise

dist.enable_validation(False)

Listing All Available Distributions

import pyro.distributions as dist

# Print all exported names
for name in sorted(dist.__all__):
    print(name)

Related Pages

• RelaxedStraightThrough -- Straight-through gradient variants of relaxed discrete distributions
• SineSkewed -- Sine-skewed torus distributions for directional statistics
• SoftLaplace -- Smooth Laplace-like distribution
• Stable -- Levy alpha-stable distribution family
• TransformModule -- Base class for learnable distribution transforms
• Unit -- Trivial distribution for factor statements