How pyro-ppl/numpyro works
NumPyro is a JAX-powered probabilistic programming library. Users define models with primitives and distributions, then run inference—especially MCMC—or variational inference to obtain posterior samples or fitted approximations; diagnostics help assess results. Effect handlers and inference utilities provide an internal layer for interpreting model execution and preparing densities. The README documents the main workflow and capabilities; the source index supports several internal links, while unsampled areas are represented only where documented.
Main entry points: __init__.py Public API · distribution.py Distributions · primitives.py Model primitives · mcmc.py MCMC runner
Components 22
Modeling
- Model primitives
numpyro/primitives.py - Distributions
numpyro/distributions/distribution.py - Continuous families
numpyro/distributions/continuous.py - Discrete families
numpyro/distributions/discrete.py - Support constraints
numpyro/distributions/constraints.py - Bijective transforms
numpyro/distributions/transforms.py
Inference
- MCMC runner
numpyro/infer/mcmc.py - HMC and NUTS
numpyro/infer/hmc.py - Variational inference
numpyro/infer/svi.py - ELBO objectives
numpyro/infer/elbo.py - Automatic guides
numpyro/infer/autoguide.py - Inference utilities
numpyro/infer/util.py - Model inspection
numpyro/infer/inspect.py - Sampling diagnostics
numpyro/diagnostics.py
Runtime Core
- Public API
numpyro/__init__.py - Effect handlers
numpyro/handlers.py - Shared types
numpyro/_typing.py
Extensions
- Discrete enumeration
numpyro/contrib/funsor/enum_messenger.py - TFP distributions
numpyro/contrib/tfp/distributions.py - Stein variational inference
numpyro/contrib/einstein/steinvi.py - HSGP approximation
numpyro/contrib/hsgp/approximation.py
Other components
- Model author
Connections 25
- Model author → Public API: uses
- Public API → Model primitives: exports
numpyro/__init__.py - Public API → Effect handlers: exports
numpyro/__init__.py - Public API → Distributions: exports
numpyro/__init__.py - Public API → MCMC runner: exports inference
numpyro/__init__.py - Model primitives → Effect handlers: dispatches messages
numpyro/primitives.py - Model primitives → Distributions: samples from
numpyro/primitives.py - Distributions → Support constraints: uses supports
numpyro/distributions/distribution.py - Distributions → Bijective transforms: uses transforms
numpyro/distributions/distribution.py - Model author → Model primitives: defines model
- Model author → Distributions: chooses distributions
- MCMC runner → HMC and NUTS: runs kernel
numpyro/infer/hmc.py - Model author → MCMC runner: runs inference
- MCMC runner → Sampling diagnostics: reports samples
- Variational inference → ELBO objectives: optimizes objective
- Variational inference → Automatic guides: uses guides
numpyro/infer/autoguide.py - ELBO objectives → Inference utilities: evaluates model
numpyro/infer/elbo.py - HMC and NUTS → Inference utilities: computes potential
numpyro/infer/hmc.py - Inference utilities → Effect handlers: conditions model
numpyro/infer/util.py - Inference utilities → Support constraints: constrains values
numpyro/infer/util.py - Public API → Model inspection: exports inspection
numpyro/__init__.py - Effect handlers → Discrete enumeration: supports enumeration
numpyro/handlers.py - Model author → TFP distributions: can use distributions
- Model author → Stein variational inference: may choose extension
- Model author → HSGP approximation: may choose extension