Tune.KNN (Tune v0.5.0)

Copy Markdown View Source

A program that selects nearby training examples for each input.

For every call it ranks the declared pool by similarity to the input, takes the k nearest rows (pool order breaks ties), runs the inner program on each neighbor under the explicit :teacher role to bootstrap demonstrations from successful trajectories, adds labeled demonstrations when configured, and then runs the inner program with those demonstrations under the active role. Every teacher call is a managed call in the same execution, so the run needs a teacher configuration and an explicit :max_model_calls.

The pool is a declared JSON parameter named "knn_pool", so it is saved in artifacts. The vectorizer is application code and is never saved. The default vectorizer is a bag of lowercase words over the input values.

Summary

Functions

Builds a KNN program around an inner program.

Returns the JSON schema used for the pool parameter.

Types

t()

@type t() :: %Tune.KNN{
  inner: term(),
  k: term(),
  max_bootstrapped_demos: term(),
  max_labeled_demos: term(),
  pool: term(),
  vectorizer: term()
}

Functions

new(inner, opts)

@spec new(Tune.Program.t(), keyword()) :: {:ok, t()} | {:error, Exception.t()}

Builds a KNN program around an inner program.

Options: :pool (training rows), :k, :max_bootstrapped_demos, :max_labeled_demos, and :vectorizer (a function of an input map returning a list of numbers, or a map of term to weight).

pool_schema(signature)

@spec pool_schema(Tune.Signature.t()) :: Zoi.schema()

Returns the JSON schema used for the pool parameter.