de.jungblut.classification.eval

Class EvaluationListener<A extends Classifier>

java.lang.Object

de.jungblut.classification.eval.EvaluationListener<A>

Type Parameters:

A - the type of the classifier.

All Implemented Interfaces:

IterationCompletionListener

public class EvaluationListener<A extends Classifier>
extends Object
implements IterationCompletionListener

The evaluation listener is majorly used to track the overfitting of a classifier while training. This is usually hooked into the Minimizer of choice and will be triggered at a configurable interval of iterations (through setRunInterval(int)). This class is designed to be subclasses and enhanced with other print statements or functionality to save the best performing parameters.

Author:

thomas.jungblut

Field Summary

Fields

Modifier and Type	Field and Description
protected WeightMapper<A>	mapper
protected int	runInterval
protected EvaluationSplit	split

Constructor Summary

Constructors

Constructor and Description
EvaluationListener(WeightMapper<A> mapper, EvaluationSplit split) Initializes this listener.
EvaluationListener(WeightMapper<A> mapper, EvaluationSplit split, int runInterval) Initializes this listener.

Method Summary

Modifier and Type	Method and Description
void	onIterationFinished(int iteration, double cost, de.jungblut.math.DoubleVector currentWeights) This callback is called from a AbstractMinimizer when a iteration of a minimization objective is finished.
protected void	onResult(int iteration, double cost, Evaluator.EvaluationResult trainEval, Evaluator.EvaluationResult testEval) Will be called on a result of the evaluation.
void	setRunInterval(int runInterval) Sets the run intervall of this listener.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

split

protected final EvaluationSplit split

mapper

protected final WeightMapper<A extends Classifier> mapper

runInterval

protected int runInterval

Constructor Detail

EvaluationListener

public EvaluationListener(WeightMapper<A> mapper,
                          EvaluationSplit split)

Initializes this listener.

Parameters:

mapper - the mapper that converts the DoubleVector from the minimizable CostFunction to a classifier.

split - the train/test split.

EvaluationListener

public EvaluationListener(WeightMapper<A> mapper,
                          EvaluationSplit split,
                          int runInterval)

Initializes this listener.

Parameters:

mapper - the mapper that converts the DoubleVector from the minimizable CostFunction to a classifier.

split - the train/test split.

runInterval - test interval.

Method Detail

onIterationFinished

public void onIterationFinished(int iteration,
                                double cost,
                                de.jungblut.math.DoubleVector currentWeights)

Description copied from interface: IterationCompletionListener

This callback is called from a AbstractMinimizer when a iteration of a minimization objective is finished.

Specified by:

onIterationFinished in interface IterationCompletionListener

Parameters:

iteration - the number of the current iteration.

cost - the cost at the current iteration.

currentWeights - the current optimal weights.

setRunInterval

public final void setRunInterval(int runInterval)

Sets the run intervall of this listener. For example: if set to 5, the evaluator will run only every five iterations.

onResult

protected void onResult(int iteration,
                        double cost,
                        Evaluator.EvaluationResult trainEval,
                        Evaluator.EvaluationResult testEval)

Will be called on a result of the evaluation. This method does nothing, is designed to be overridden though.

Parameters:

iteration - the current number of iteration.

cost - the identified cost of the costfunction.

trainEval - the evaluation on the trainingset.

testEval - the evaluation on the testset.