org.apache.ignite.ml.composition.boosting

Class GDBTrainer

java.lang.Object

org.apache.ignite.ml.trainers.DatasetTrainer<ModelsComposition,Double>

org.apache.ignite.ml.composition.boosting.GDBTrainer

Direct Known Subclasses:

GDBBinaryClassifierTrainer, GDBRegressionTrainer

public abstract class GDBTrainer
extends DatasetTrainer<ModelsComposition,Double>

Abstract Gradient Boosting trainer. It implements gradient descent in functional space using user-selected regressor in child class. Each learning iteration the trainer evaluate gradient of error-function and fit regression model to it. After learning step the model is used in models composition of regressions with weight equal to gradient descent step. These classes can be used as regressor trainers: DecisionTreeRegressionTrainer, KNNRegressionTrainer, LinearRegressionLSQRTrainer, RandomForestRegressionTrainer, LinearRegressionSGDTrainer. But in practice Decision Trees is most used regressors (see: DecisionTreeRegressionTrainer).

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	GDBTrainer.GDBModel GDB model.

Nested classes/interfaces inherited from class org.apache.ignite.ml.trainers.DatasetTrainer

DatasetTrainer.EmptyDatasetException

Field Summary

Fields

Modifier and Type	Field and Description
protected ConvergenceCheckerFactory	checkConvergenceStgyFactory Check convergence strategy factory.
protected Loss	loss Loss function.

Fields inherited from class org.apache.ignite.ml.trainers.DatasetTrainer

envBuilder, environment

Constructor Summary

Constructors

Constructor and Description
GDBTrainer(double gradStepSize, Integer cntOfIterations, Loss loss) Constructs GDBTrainer instance.

Method Summary

Modifier and Type	Method and Description
protected abstract @NotNull DatasetTrainer<? extends IgniteModel<Vector,Double>,Double>	buildBaseModelTrainer() Returns regressor model trainer for one step of GDB.
protected <V,K,C extends Serializable> IgniteBiTuple<Double,Long>	computeInitialValue(LearningEnvironmentBuilder envBuilder, DatasetBuilder<K,V> builder, Preprocessor<K,V> preprocessor) Compute mean value of label as first approximation.
protected abstract double	externalLabelToInternal(double lbl) Maps external representation of label to internal.
<K,V> ModelsComposition	fitWithInitializedDeployingContext(DatasetBuilder<K,V> datasetBuilder, Preprocessor<K,V> preprocessor) Trains model based on the specified data.
protected GDBLearningStrategy	getLearningStrategy() Returns learning strategy.
protected abstract double	internalLabelToExternal(double lbl) Maps internal representation of label to external.
boolean	isUpdateable(ModelsComposition mdl)
protected abstract <V,K> boolean	learnLabels(DatasetBuilder<K,V> builder, Preprocessor<K,V> preprocessor) Defines unique labels in dataset if need (useful in case of classification).
protected <K,V> ModelsComposition	updateModel(ModelsComposition mdl, DatasetBuilder<K,V> datasetBuilder, Preprocessor<K,V> preprocessor) Trains new model taken previous one as a first approximation.
GDBTrainer	withCheckConvergenceStgyFactory(ConvergenceCheckerFactory factory) Sets CheckConvergenceStgyFactory.
GDBTrainer	withEnvironmentBuilder(LearningEnvironmentBuilder envBuilder) Changes learning Environment.

Methods inherited from class org.apache.ignite.ml.trainers.DatasetTrainer

fit, fit, fit, fit, fit, fit, getLastTrainedModelOrThrowEmptyDatasetException, identityTrainer, learningEnvironment, update, update, update, update, update, withConvertedLabels

Methods inherited from class java.lang.Object

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

Field Detail

loss

protected final Loss loss

Loss function.

checkConvergenceStgyFactory

protected ConvergenceCheckerFactory checkConvergenceStgyFactory

Check convergence strategy factory.

Constructor Detail

GDBTrainer

public GDBTrainer(double gradStepSize,
                  Integer cntOfIterations,
                  Loss loss)

Constructs GDBTrainer instance.

Parameters:

gradStepSize - Grad step size.

cntOfIterations - Count of learning iterations.

loss - Gradient of loss function. First argument is sample size, second argument is valid answer third argument is current model prediction.

Method Detail

fitWithInitializedDeployingContext

public <K,V> ModelsComposition fitWithInitializedDeployingContext(DatasetBuilder<K,V> datasetBuilder,
                                                                  Preprocessor<K,V> preprocessor)

Trains model based on the specified data.

Specified by:

fitWithInitializedDeployingContext in class DatasetTrainer<ModelsComposition,Double>

Type Parameters:

K - Type of a key in upstream data.

V - Type of a value in upstream data.

Parameters:

datasetBuilder - Dataset builder.

preprocessor - Extractor of UpstreamEntry into LabeledVector.

Returns:

Model.

updateModel

protected <K,V> ModelsComposition updateModel(ModelsComposition mdl,
                                              DatasetBuilder<K,V> datasetBuilder,
                                              Preprocessor<K,V> preprocessor)

Trains new model taken previous one as a first approximation.

Specified by:

updateModel in class DatasetTrainer<ModelsComposition,Double>

Type Parameters:

K - Type of a key in upstream data.

V - Type of a value in upstream data.

Parameters:

mdl - Learned model.

datasetBuilder - Dataset builder.

preprocessor - Extractor of UpstreamEntry into LabeledVector.

Returns:

Updated model.

isUpdateable

public boolean isUpdateable(ModelsComposition mdl)

Specified by:

isUpdateable in class DatasetTrainer<ModelsComposition,Double>

Parameters:

mdl - Model.

Returns:

True if current critical for training parameters correspond to parameters from last training.

withEnvironmentBuilder

public GDBTrainer withEnvironmentBuilder(LearningEnvironmentBuilder envBuilder)

Changes learning Environment.

Overrides:

withEnvironmentBuilder in class DatasetTrainer<ModelsComposition,Double>

Parameters:

envBuilder - Learning environment builder.

learnLabels

protected abstract <V,K> boolean learnLabels(DatasetBuilder<K,V> builder,
                                             Preprocessor<K,V> preprocessor)

Defines unique labels in dataset if need (useful in case of classification).

Parameters:

builder - Dataset builder.

preprocessor - Upstream preprocessor.

Returns:

True if labels learning was successful.

buildBaseModelTrainer

@NotNull
protected abstract @NotNull DatasetTrainer<? extends IgniteModel<Vector,Double>,Double> buildBaseModelTrainer()

Returns regressor model trainer for one step of GDB.

externalLabelToInternal

protected abstract double externalLabelToInternal(double lbl)

Maps external representation of label to internal.

Parameters:

lbl - Label value.

internalLabelToExternal

protected abstract double internalLabelToExternal(double lbl)

Maps internal representation of label to external.

Parameters:

lbl - Label value.

computeInitialValue

protected <V,K,C extends Serializable> IgniteBiTuple<Double,Long> computeInitialValue(LearningEnvironmentBuilder envBuilder,
                                                                                      DatasetBuilder<K,V> builder,
                                                                                      Preprocessor<K,V> preprocessor)

Compute mean value of label as first approximation.

Parameters:

builder - Dataset builder.

envBuilder - Learning environment builder.

preprocessor - Vectorizer.

withCheckConvergenceStgyFactory

public GDBTrainer withCheckConvergenceStgyFactory(ConvergenceCheckerFactory factory)

Sets CheckConvergenceStgyFactory.

Parameters:

factory - Factory.

Returns:

Trainer.

getLearningStrategy

protected GDBLearningStrategy getLearningStrategy()

Returns learning strategy.

Returns:

Learning strategy.