Package ai.djl.training.optimizer

Class RmsProp

java.lang.Object

ai.djl.training.optimizer.Optimizer

ai.djl.training.optimizer.RmsProp

public class RmsProp
extends Optimizer

The RMSProp Optimizer.

Two versions of RMSProp are implemented.

If `centered = False`, the algorithm described in http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf by Tieleman and Hinton, 2012 is used.

If `centered = True`, the algorithm described in http://arxiv.org/pdf/1308.0850v5.pdf (38)-(45) by Alex Graves, 2013 is used instead.

Default version is `centered = False`.

If `centered = False`:

RMSProp updates the weights using:

\( var = rho * var + (1 - rho) * grad^2 \)
\( weight -= learning_rate * (sqrt(v) + epsilon) \)

If `centered = True`: \( mean = rho * mean + (1 - rho) * grad \)
\( var = rho * var + (1 - rho) * grad^2 \)
\( mom = mom^2 - lr * grad / sqrt(var - mean^2) + epsilon \)
\( weight = mean / (sqrt(var) + epsilon) \)

Grad represents the gradient, mean and var are the 1st and 2nd order moment estimates (mean and variance), and mom is the momentum.

See Also:

• The D2L chapter on RMSProp

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static final class	RmsProp.Builder	The Builder to construct an RmsProp object.

Nested classes/interfaces inherited from class ai.djl.training.optimizer.Optimizer

Optimizer.OptimizerBuilder<T extends Optimizer.OptimizerBuilder>

Field Summary

Fields inherited from class ai.djl.training.optimizer.Optimizer

clipGrad, rescaleGrad

Constructor Summary

Constructors

Modifier	Constructor	Description
protected	RmsProp(RmsProp.Builder builder)	Creates a new instance of RMSProp optimizer.

Method Summary

Modifier and Type	Method	Description
static RmsProp.Builder	builder()	Creates a builder to build a RMSProp.
void	update(String parameterId, NDArray weight, NDArray grad)	Updates the parameters according to the gradients.

Methods inherited from class ai.djl.training.optimizer.Optimizer

adadelta, adagrad, adam, adamW, getWeightDecay, nag, rmsprop, sgd, updateCount, withDefaultState

Methods inherited from class java.lang.Object

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

Constructor Details

RmsProp

protected RmsProp(RmsProp.Builder builder)

Creates a new instance of RMSProp optimizer.

Parameters:

builder - the builder to create a new instance of Adam optimizer

Method Details

update

public void update(String parameterId,
 NDArray weight,
 NDArray grad)

Updates the parameters according to the gradients.

Specified by:

update in class Optimizer

Parameters:

parameterId - the parameter to be updated

weight - the weights of the parameter

grad - the gradients

builder

public static RmsProp.Builder builder()

Creates a builder to build a RMSProp.

Returns:

a new builder