Probabilistic Programming API

This page is a flat index of the public exports of Reactant.ProbProg, Reactant's probabilistic-programming module. Names are qualified as ProbProg.<name>; using Reactant: ProbProg unqualifies them.

For walkthroughs and runnable end-to-end programs, follow the Probabilistic Programming tutorials:

• Overview — Bayesian linear and logistic regression.

• Interface overview — quick tabular summary of the public surface.

• Sampling and distributions — sample, built-in distributions, custom samplers, supports.

• Traces and constrained inference — Trace, Address, Constraint, Selection, simulate and generate.

• Trace representation — how Reactant carries a trace as a flat position-vector tensor during compilation, and the helpers that expand it back into a tree-shaped Trace.

• MCMC: MH, HMC, NUTS — mh, mcmc, mcmc_logpdf.

• Running and resuming chains — run_chain, MCMCState, checkpointing, mcmc_summary.

The conceptual vocabulary (traces, choice maps, selections, generative functions) is the same as in Gen.jl; readers familiar with Gen can map Constraint onto Gen's ChoiceMap, select onto Gen.select, and so on.

Types

Symbol	Signature	See
Trace	mutable struct Trace; choices, retval, weight, subtraces; end	Traces and constrained inference
Address	Address(syms::Symbol...)	Traces and constrained inference
Selection	const Selection = OrderedSet{Address}; built via select	Traces and constrained inference
Constraint	Constraint(pairs::Pair...)	Traces and constrained inference
MCMCState	mutable struct MCMCState; position, gradient, potential_energy, step_size, inverse_mass_matrix, rng; end	Running and resuming chains

Distributions

ProbProg.sample(rng, dist; symbol=:x) works on any Distribution subtype. Built-in distributions (Normal, Exponential, LogNormal, Bernoulli) and the recipe for defining custom ones are described in the Sampling and distributions tutorial.

Modeling

Symbol	Signature	See
sample	sample(rng, dist; symbol)  /  sample(rng, f, args...; symbol, logpdf, support, bounds)	Sampling and distributions
untraced_call	untraced_call(rng, f, args...)	Sampling and distributions
simulate	simulate(rng, f, args...) -> (trace_tensor, weight, retval)	Traces and constrained inference
simulate_	simulate_(rng, f, args...) -> (trace::Trace, weight)	Traces and constrained inference
generate	generate(rng, constraint_tensor, f, args...; constrained_addresses)	Traces and constrained inference
generate_	generate_(rng, constraint, f, args...) -> (trace::Trace, weight)	Traces and constrained inference

Inference

Symbol	Signature	See
mh	mh(rng, trace, weight, f, args...; selection)	MCMC: MH, HMC, NUTS
mcmc	mcmc(rng, original_trace, f, args...; selection, algorithm, step_size, inverse_mass_matrix, num_warmup, num_samples, ...)  /  mcmc(state::MCMCState, ...)	MCMC: MH, HMC, NUTS
mcmc_logpdf	mcmc_logpdf(rng, logdensity_fn, initial_position, args...; algorithm, step_size, inverse_mass_matrix, ...)  /  mcmc_logpdf(state::MCMCState, ...)	MCMC: MH, HMC, NUTS
run_chain	run_chain(rng, logpdf_fn, initial_position, args...; num_warmup, num_samples, chunk_size, ...)  /  run_chain(state::MCMCState, ...)	Running and resuming chains

NUTS currently requires an explicit step_size; the default step_size = nothing is rejected by the pass implementation (find_reasonable_step_size is not yet implemented). Pass Reactant.ConcreteRNumber(0.1) (or similar) and rely on dual averaging when adapt_step_size = true.

State persistence

Symbol	Signature	See
save_state	save_state(filename::String, state::MCMCState)	Running and resuming chains
load_state	load_state(filename::String) -> MCMCState	Running and resuming chains

Utilities

Symbol	Signature	See
select	select(addrs::Address...) -> Selection	Traces and constrained inference
get_choices	get_choices(trace::Trace) -> Dict{Symbol,Any}	Traces and constrained inference
with_trace	with_trace(f) -> (f_result, tt) (installs the Impulse tracing context for the duration of f(); collects the layout metadata unflatten_trace needs)	Traces and constrained inference
unflatten_trace	unflatten_trace(trace_tensor, weight, entries, retval) -> Trace	Traces and constrained inference
filter_entries_by_selection	filter_entries_by_selection(entries, selection)	MCMC: MH, HMC, NUTS
extract_addresses	extract_addresses(constraint::Constraint) -> Set{Address}	Traces and constrained inference
mcmc_summary	mcmc_summary(samples; names)  /  mcmc_summary(trace::Trace)	Running and resuming chains