org.nd4j.linalg.activations.impl

Class ActivationPReLU

java.lang.Object

org.nd4j.linalg.activations.BaseActivationFunction

org.nd4j.linalg.activations.impl.ActivationPReLU

All Implemented Interfaces:

Serializable, IActivation

public class ActivationPReLU
extends BaseActivationFunction

/** Parametrized Rectified Linear Unit (PReLU) f(x) = alpha * x for x < 0, f(x) = x for x >= 0 alpha has the same shape as x and is a learned parameter.

Author:

Max Pumperla

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
ActivationPReLU(INDArray alpha)
ActivationPReLU(INDArray alpha, long[] sharedAxes)

Method Summary

Modifier and Type	Method and Description
org.nd4j.linalg.primitives.Pair<INDArray,INDArray>	backprop(INDArray in, INDArray epsilon) Backpropagate the errors through the activation function, given input z and epsilon dL/da. Returns 2 INDArrays: (a) The gradient dL/dz, calculated from dL/da, and (b) The parameter gradients dL/dW, where w is the weights in the activation function.
INDArray	getActivation(INDArray in, boolean training) Carry out activation function on the input array (usually known as 'preOut' or 'z') Implementations must overwrite "in", transform in place and return "in" Can support separate behaviour during test
String	toString()

Methods inherited from class org.nd4j.linalg.activations.BaseActivationFunction

assertShape, getGradientViewArray, getParametersViewArray, numParams, setGradientViewArray, setParametersViewArray

Methods inherited from class java.lang.Object

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

Constructor Detail

ActivationPReLU

public ActivationPReLU(INDArray alpha)

ActivationPReLU

public ActivationPReLU(INDArray alpha,
                       long[] sharedAxes)

Method Detail

getActivation

public INDArray getActivation(INDArray in,
                              boolean training)

Description copied from interface: IActivation

Carry out activation function on the input array (usually known as 'preOut' or 'z') Implementations must overwrite "in", transform in place and return "in" Can support separate behaviour during test

Returns:

transformed activation

backprop

public org.nd4j.linalg.primitives.Pair<INDArray,INDArray> backprop(INDArray in,
                                                                   INDArray epsilon)

Description copied from interface: IActivation

Backpropagate the errors through the activation function, given input z and epsilon dL/da.
Returns 2 INDArrays:
(a) The gradient dL/dz, calculated from dL/da, and
(b) The parameter gradients dL/dW, where w is the weights in the activation function. For activation functions with no gradients, this will be null.

Parameters:

in - Input, before applying the activation function (z, or 'preOut')

epsilon - Gradient to be backpropagated: dL/da, where L is the loss function

Returns:

dL/dz and dL/dW, for weights w (null if activation function has no weights)

toString

public String toString()

Overrides:

toString in class Object