GitHub

Input arguments annotations

Background: tracing rule inputs

During inference, every message flowing along an edge is computed by a message update rule. InputArgumentsAnnotations records what went into each rule call — the MessageMapping (which node and interface the rule was for), the incoming messages, the incoming marginals, and the result distribution — and propagates that record through subsequent message products.

This is useful for debugging and for implementing callbacks that need to inspect the full provenance of a message: rather than re-running or re-examining the model structure, the record travels with the message itself.

What gets stored

After each rule execution a RuleInputArgumentsRecord is written into the message's annotation dict under the :rule_input_arguments key. When two messages are multiplied, their records are merged into a ProductInputArgumentsRecord that contains all contributing records as a flat list, regardless of how deeply nested the products were.

Reading input arguments from a message

using ReactiveMP

# ann is the AnnotationDict of some message
record = get_rule_input_arguments(ann)

if record isa RuleInputArgumentsRecord
    println("single rule: ", record.mapping)
    println("messages:  ", record.messages)
    println("marginals: ", record.marginals)
    println("result:    ", record.result)
elseif record isa ProductInputArgumentsRecord
    for r in record.mappings
        println("contributed rule: ", r.mapping)
    end
end

API

ReactiveMP.RuleInputArgumentsRecord — Type
RuleInputArgumentsRecord

Stores the inputs and result of a single message update rule execution: the MessageMapping, the incoming messages tuple, the incoming marginals tuple, and the computed result distribution.

source