Latu.ML.Estimator (latu_ml v0.2.0)

An unfitted estimator: inert data, no session, no server-side anything.

Latu.ML.fit/2 is what turns one into a Latu.ML.Model, and the frame it is fitted on is where the session comes from.

Summary

Functions

new(class, opts \\ [])

An estimator the registry does not know, by its JVM class name.

Types

t()

@type t() :: %Latu.ML.Estimator{
  class: String.t(),
  known: [Latu.ML.Param.t()],
  model_candidates: [String.t()],
  model_class: String.t() | nil,
  params: [Latu.ML.Plan.param()],
  uid: String.t()
}

Functions

new(class, opts \\ [])

@spec new(String.t(), keyword()) :: t()

An estimator the registry does not know, by its JVM class name.

The escape hatch for anything on the server's classpath that MLlib did not put there. Nothing checks the param names — there is no table to check them against — so the server's error is the only one you get, and model_class is yours to supply if you want the model's attributes to resolve later.

Options

:params — the params to send, already in wire form: {name, type, value} per param, where name is Spark's own spelling. Defaults to [].
:model_class — the JVM class of the model a fit produces. Defaults to nil, which leaves Latu.ML.Model's :class unset.
:uid — the uid to send. Defaults to a generated one in PySpark's shape.

Example

Latu.ML.Estimator.new("ml.dmlc.xgboost4j.scala.spark.XGBoostClassifier",
  params: [{"numRound", :int, 100}, {"max_depth", :int, 6}])

A param's name is whatever the library declared it, and that is not always one convention: XGBoost4J spells its own params camelCase and the native booster's snake_case, and the server resolves each exactly as written.