Skip to content

README.md

Latest commit

 

History

History
76 lines (53 loc) · 2.57 KB

README.md

File metadata and controls

76 lines (53 loc) · 2.57 KB

MLJTestIntegration.jl

Package for applying integration tests to models implementing the MLJ model interface.

For to apply minimal tests for an implementations of the MLJ model interface, use MLJTestInterface.jl instead.

However, implementations are now encouraged to include the comprehensive tests provided in this package instead.

Lifecycle:Experimental Build Status Coverage

Installation

using Pkg
Pkg.add("MLJTestIntegration")

Usage

This package provides a method for testing a collection of models (more, specifically model types) using the specified training data:

MLJTestIntegration.test(models, data...; mod=Main, level=2, throw=false, verbosity=1)
    -> failures, summary

For pure Julia models, level=4 (the maximimum) is recommended. Otherwise level=3 is recommended.

For detailed documentation, run using MLJTestIntegration; @doc MLJTestIntegration.test.

Example

The following tests the model interface implemented by the DecisionTreeClassifier model implemented in the package MLJDecisionTreeInterface.jl.

import MLJDecisionTreeInterface
import MLJTestIntegration
using Test
X, y = MLJTestIntegration.make_multiclass()
failures, summary = MLJTestIntegration.test(
    [MLJDecisionTreeInterface.DecisionTreeClassifier, ],
    X, y,
    verbosity=0, # set to 2 when debugging
    throw=false, # set to `true` when debugging
    mod=@__MODULE__,
)
@test isempty(failures)

For models supporting tabular data X, we recommend repeating the tests with (X, y) replaced with (Tables.rowtable(X)), y) to ensure support for different table types, which new implementation sometimes fail to do.

Datasets

The following commands generate datasets of the form (X, y) suitable for integration tests:

  • MLJTestIntegration.make_binary

  • MLJTestIntegration.make_multiclass

  • MLJTestIntegration.make_regression

  • MLJTestIntegration.make_count

The binary and multiclass sets are deliberately small and will catch failures to properly track target class pools, a common gotcha in new implementations.