Monty.Frameworks.Models.EvidenceMatching.FeatureEvidence.Calculator.FeatureEvidenceCalculator behaviour (ExTbpMonty.Core v0.1.0)

Copy Markdown View Source

What a feature-evidence calculator has to be able to do.

Mirrors tbp/monty/frameworks/models/evidence_matching/feature_evidence/calculator.py::FeatureEvidenceCalculator, which is a Protocol: any object with a calculate of this shape can be used as one. In Elixir that is a behaviour, and the hypotheses updater takes the calculator as a module, so an experiment can substitute its own — Monty's dependency injection, kept.

Shapes

channel_feature_array is one row per node, one column per stored feature value; channel_feature_order names the features in the order their columns appear. A multi-valued feature (hsv with three values) occupies three columns, and how many is read from the query features' values — that is the only place the length is written down, exactly as in Python.

pose_vectors and pose_fully_defined are not feature evidence: they are compared as poses elsewhere and occupy no columns here.

Summary

Types

feature_values()
row()
tolerance()

Callbacks

calculate( channel_feature_array, channel_feature_order, channel_feature_weights, channel_query_features, channel_tolerances )

Evidence in [0, 1] for every node in the graph, or nil where a node's stored values are missing.

Types

feature_values()

@type feature_values() :: number() | [number()]

row()

@type row() :: [float() | nil]

tolerance()

@type tolerance() :: number() | [number()]

Callbacks

calculate( channel_feature_array, channel_feature_order, channel_feature_weights, channel_query_features, channel_tolerances )

@callback calculate(
  channel_feature_array :: [row()],
  channel_feature_order :: [String.t()],
  channel_feature_weights :: %{optional(String.t()) => number()},
  channel_query_features :: %{optional(String.t()) => feature_values()},
  channel_tolerances :: %{optional(String.t()) => tolerance()}
) :: [float() | nil]

Evidence in [0, 1] for every node in the graph, or nil where a node's stored values are missing.

nil is this port's NaN: a value that is not a number is not a score, and callers must decide what a node without data means rather than have a comparison quietly go one way.