Class NetworkTrainer

java.lang.Object

org.ojalgo.ann.WrappedANN

org.ojalgo.ann.NetworkTrainer

All Implemented Interfaces:

Supplier<ArtificialNeuralNetwork>

public final class NetworkTrainer
extends WrappedANN

An Artificial Neural Network (ANN) builder/trainer.

Field Summary

Fields

Modifier and Type	Field	Description
private final TrainingConfiguration	myConfiguration
private final PhysicalStore<Double>[]	myGradients

Constructor Summary

Constructors

Constructor	Description
NetworkTrainer(ArtificialNeuralNetwork network, int batchSize)

Method Summary

Modifier and Type	Method	Description
NetworkTrainer	activator(int layer, ArtificialNeuralNetwork.Activator activator)	Deprecated. Use NetworkBuilder and NetworkBuilder.layer(int, Activator) instead.
NetworkTrainer	activators(ArtificialNeuralNetwork.Activator activator)	Deprecated. Use NetworkBuilder and NetworkBuilder.layer(int, Activator) instead.
NetworkTrainer	activators(ArtificialNeuralNetwork.Activator... activators)	Deprecated. Use NetworkBuilder and NetworkBuilder.layer(int, Activator) instead.
NetworkTrainer	bias(int layer, int output, double bias)
NetworkTrainer	dropouts()
boolean	equals(Object obj)
NetworkTrainer	error(ArtificialNeuralNetwork.Error error)
(package private) double	error(Access1D<?> target, Access1D<?> current)
int	hashCode()
NetworkTrainer	lasso(double factor)	L1 lasso regularisation
DataBatch	newOutputBatch()
NetworkTrainer	rate(double rate)
NetworkTrainer	ridge(double factor)	L2 ridge regularisation
Structure2D[]	structure()
String	toString()
void	train(Iterable<? extends Access1D<Double>> givenInputs, Iterable<? extends Access1D<Double>> targetOutputs)	Deprecated. Just use train(Access1D, Access1D) instead
void	train(Access1D<Double> givenInput, Access1D<Double> targetOutput)	The arguments are typed as Access1D but it's probably best to think of (create) them as something 2D where the number of rows should match the batch size and the number of columns the number of inputs and outputs respectively.
NetworkTrainer	weight(int layer, int input, int output, double weight)

Methods inherited from class WrappedANN

adjust, depth, get, getActivator, getBatchSize, getBias, getInput, getInput, getOutput, getOutput, getOutputActivator, getWeight, getWeights, invoke, newInputBatch, randomise, setActivator, setBias, setWeight

Methods inherited from class Object

clone, finalize, getClass, notify, notifyAll, wait, wait, wait

Field Details

myConfiguration

private final TrainingConfiguration myConfiguration

myGradients

private final PhysicalStore<Double>[] myGradients

Constructor Details

NetworkTrainer

NetworkTrainer(ArtificialNeuralNetwork network,
 int batchSize)

Method Details

activator

@Deprecated
public NetworkTrainer activator(int layer,
 ArtificialNeuralNetwork.Activator activator)

Deprecated.

Use NetworkBuilder and NetworkBuilder.layer(int, Activator) instead.

Parameters:

layer - 0-based index among the calculation layers (excluding the input layer)

activator - The activator function to use

activators

@Deprecated
public NetworkTrainer activators(ArtificialNeuralNetwork.Activator activator)

Deprecated.

Use NetworkBuilder and NetworkBuilder.layer(int, Activator) instead.

activators

@Deprecated
public NetworkTrainer activators(ArtificialNeuralNetwork.Activator... activators)

Deprecated.

Use NetworkBuilder and NetworkBuilder.layer(int, Activator) instead.

bias

public NetworkTrainer bias(int layer,
 int output,
 double bias)

dropouts

public NetworkTrainer dropouts()

equals

public boolean equals(Object obj)

Overrides:

equals in class WrappedANN

error

public NetworkTrainer error(ArtificialNeuralNetwork.Error error)

hashCode

public int hashCode()

Overrides:

hashCode in class WrappedANN

lasso

public NetworkTrainer lasso(double factor)

L1 lasso regularisation

newOutputBatch

public DataBatch newOutputBatch()

Overrides:

newOutputBatch in class WrappedANN

See Also:

• WrappedANN.newInputBatch()

rate

public NetworkTrainer rate(double rate)

ridge

public NetworkTrainer ridge(double factor)

L2 ridge regularisation

structure

public Structure2D[] structure()

Overrides:

structure in class WrappedANN

toString

public String toString()

Overrides:

toString in class Object

train

public void train(Access1D<Double> givenInput,
 Access1D<Double> targetOutput)

The arguments are typed as Access1D but it's probably best to think of (create) them as something 2D where the number of rows should match the batch size and the number of columns the number of inputs and outputs respectively. When the batch size is 1 then the arguments can actually be 1D.

Parameters:

givenInput - One or more input examples, depending on the batch size

targetOutput - One or more, matching, output targets

train

@Deprecated
public void train(Iterable<? extends Access1D<Double>> givenInputs,
 Iterable<? extends Access1D<Double>> targetOutputs)

Deprecated.

Just use train(Access1D, Access1D) instead

Note that the required Iterable:s can be obtained from calling Access2D.rows() or Access2D.columns() on anything "2D".

weight

public NetworkTrainer weight(int layer,
 int input,
 int output,
 double weight)

error

double error(Access1D<?> target,
 Access1D<?> current)