de.jungblut.online.minimizer

Class StochasticGradientDescent

java.lang.Object

de.jungblut.online.minimizer.StochasticGradientDescent

All Implemented Interfaces:

StochasticMinimizer

public class StochasticGradientDescent
extends Object
implements StochasticMinimizer

Stochastic gradient descent. This class is designed to work on a parallel stream and do stochastic updates to a parameter set.

Author:

thomas.jungblut

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	StochasticGradientDescent.StochasticGradientDescentBuilder

Method Summary

Modifier and Type	Method and Description
void	computeMomentum()
de.jungblut.math.DoubleVector	minimize(de.jungblut.math.DoubleVector start, java.util.function.Supplier<java.util.stream.Stream<de.jungblut.online.ml.FeatureOutcomePair>> streamSupplier, StochasticCostFunction costFunction, int numPasses, boolean verbose) Minimizes the given stochastic cost function on the supplied streams for the given amount of passes over the data.
void	setIterationCallback(IterationFinishedCallback iterationCallback)
void	setPassCallback(PassFinishedCallback passCallback)
void	setValidationCallback(ValidationFinishedCallback validationCallback)
CostWeightTuple	updateWeights(de.jungblut.math.minimize.CostGradientTuple observed)

Methods inherited from class java.lang.Object

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

Method Detail

minimize

public de.jungblut.math.DoubleVector minimize(de.jungblut.math.DoubleVector start,
                                              java.util.function.Supplier<java.util.stream.Stream<de.jungblut.online.ml.FeatureOutcomePair>> streamSupplier,
                                              StochasticCostFunction costFunction,
                                              int numPasses,
                                              boolean verbose)

Description copied from interface: StochasticMinimizer

Minimizes the given stochastic cost function on the supplied streams for the given amount of passes over the data.

Specified by:

minimize in interface StochasticMinimizer

Parameters:

start - the start parameters.

streamSupplier - the supplier for the data training streams.

costFunction - the cost function to minimize.

numPasses - the number of passes over the streams.

verbose - true if progress should be printed to the log.

Returns:

the optimized set of parameters.

computeMomentum

public void computeMomentum()

updateWeights

public CostWeightTuple updateWeights(de.jungblut.math.minimize.CostGradientTuple observed)

setIterationCallback

public void setIterationCallback(IterationFinishedCallback iterationCallback)

setValidationCallback

public void setValidationCallback(ValidationFinishedCallback validationCallback)

setPassCallback

public void setPassCallback(PassFinishedCallback passCallback)