Package org.nd4j.linalg.learning.config

Interface IUpdater

All Superinterfaces:

Cloneable, Serializable

All Known Implementing Classes:

AdaBelief, AdaDelta, AdaGrad, Adam, AdaMax, AMSGrad, Nadam, Nesterovs, NoOp, RmsProp, Sgd

public interface IUpdater
extends Serializable, Cloneable

Method Summary

Modifier and Type	Method	Description
IUpdater	clone()	Clone the updater
boolean	equals(Object updater)
double	getLearningRate(int iteration, int epoch)	Get the learning rate - if any - for the updater, at the specified iteration and epoch.
boolean	hasLearningRate()
GradientUpdater	instantiate(Map<String,INDArray> updaterState, boolean initializeStateArrays)
GradientUpdater	instantiate(INDArray viewArray, boolean initializeViewArray)	Create a new gradient updater
void	setLrAndSchedule(double lr, ISchedule lrSchedule)	Set the learning rate and schedule.
long	stateSize(long numParams)	Determine the updater state size for the given number of parameters.

Method Detail

stateSize

long stateSize(long numParams)

Determine the updater state size for the given number of parameters. Usually a integer multiple (0,1 or 2) times the number of parameters in a layer.

Parameters:

numParams - Number of parameters

Returns:

Updater state size for the given number of parameters

instantiate

GradientUpdater instantiate(INDArray viewArray,
                            boolean initializeViewArray)

Create a new gradient updater

Parameters:

viewArray - The updater state size view away

initializeViewArray - If true: initialise the updater state

Returns:

instantiate

GradientUpdater instantiate(Map<String,INDArray> updaterState,
                            boolean initializeStateArrays)

equals

boolean equals(Object updater)

Overrides:

equals in class Object

clone

IUpdater clone()

Clone the updater

getLearningRate

double getLearningRate(int iteration,
                       int epoch)

Get the learning rate - if any - for the updater, at the specified iteration and epoch. Note that if no learning rate is applicable (AdaDelta, NoOp updaters etc) then Double.NaN should be return

Parameters:

iteration - Iteration at which to get the learning rate

epoch - Epoch at which to get the learning rate

Returns:

Learning rate, or Double.NaN if no learning rate is applicable for this updater

hasLearningRate

boolean hasLearningRate()

Returns:

True if the updater has a learning rate hyperparameter, false otherwise

setLrAndSchedule

void setLrAndSchedule(double lr,
                      ISchedule lrSchedule)

Set the learning rate and schedule. Note: may throw an exception if hasLearningRate() returns false.

Parameters:

lr - Learning rate to set (typically not used if LR schedule is non-null)

lrSchedule - Learning rate schedule to set (may be null)