Imp.Experiment.Data (Imp v0.16.1)

Copy Markdown View Source

Disjoint train, selection, and untouched-test data for Imp.Experiment.check/5.

Rows are identified before any optimizer or model call. By default identity is the deterministic content digest of the complete row. Pass id: :field_name (or an arity-one function) when the dataset has a stronger source identity. Duplicate identities within or across splits are rejected. Every row is an Imp.Example that declares its inputs with Imp.with_inputs/2; a plain map, a field pair list, or an example without declared inputs raises ArgumentError here.

Summary

Types

split()
t()

Functions

manifest(data)

Returns the content-bound public description stored in experiment results.

new(opts)

Builds and validates a disjoint experiment dataset.

split(data, split)

Returns the rows for a named split.

Types

split()

@type split() :: :train | :selection | :test

t()

@type t() :: %Imp.Experiment.Data{
  digests: %{required(split()) => String.t()},
  ids: %{required(split()) => [String.t()]},
  selection: [term()],
  test: [term()],
  train: [term()]
}

Functions

manifest(data)

@spec manifest(t()) :: map()

Returns the content-bound public description stored in experiment results.

new(opts)

@spec new(keyword()) :: t()

Builds and validates a disjoint experiment dataset.

split(data, split)

@spec split(t(), split()) :: [term()]

Returns the rows for a named split.