Class MaxPoolingLayer

java.lang.Object

deepnetts.net.layers.AbstractLayer<TensorBase, TensorBase, TensorBase>

deepnetts.net.layers.MaxPoolingLayer

All Implemented Interfaces:

Backward, Forward, Layer<TensorBase>, Serializable

public final class MaxPoolingLayer
extends AbstractLayer<TensorBase, TensorBase, TensorBase>

This layer performs max pooling operation in convolutional neural network, which scales down output from previous layer by taking max outputs from small predefined filter areas.

Author:

Zoran Sevarac

See Also:

• ConvolutionalNetwork
• Serialized Form

Field Summary

Fields inherited from class AbstractLayer

activation, activationType, backwardImpl, batchMode, batchSize, biases, cudaHandles, deltaBiases, deltas, deltaWeights, depth, forwardImpl, gradients, height, inputs, learningRate, mode, momentum, networkType, nextLayer, numThreads, optimizer, optimizerType, outputs, prevDeltaBiases, prevDeltaWeights, prevLayer, randomWeightsType, regL1, regL2, threadPool, trainable, weights, width

Constructor Summary

Constructors

Constructor	Description
MaxPoolingLayer(int filterWidth, int filterHeight, int stride)	Creates a new max pooling layer with specified filter dimensions and stride.
MaxPoolingLayer(Filter filter)

Method Summary

Modifier and Type	Method	Description
void	applyWeightChanges()	Does nothing for pooling layer since it does not have weights It just propagates deltas from next layer to previous through connections that had max activation in forward pass
void	backward()	backward pass for a max(x, y) operation has a simple interpretation as only routing the gradient to the input that had the highest value in the forward pass.
void	forward()	Max pooling forward pass outputs the max value for each filter position.
int	getFilterHeight()
int	getFilterWidth()
float	getL1WeightSum()
float	getL2WeightSum()
int	getStride()
final void	init()	This method should implement layer initialization in subclasses, when a layer is added to the network (create weights, outputs, deltas, randomization etc.).
void	initTransientFields()
String	toString()

Methods inherited from class AbstractLayer

getActivation, getActivationType, getBackwardAcc, getBatchSize, getBiases, getDeltaBiases, getDeltas, getDeltaWeights, getDepth, getForwardAcc, getGradients, getHeight, getL1Regularization, getL2Regularization, getLearningRate, getMode, getMomentum, getNetworkType, getNextLayer, getNumThreads, getOptimizer, getOptimizerType, getOutputs, getPrevDeltaBiases, getPrevDeltaWeights, getPrevlayer, getWeights, getWidth, isBatchMode, isTrainable, setActivationType, setBatchMode, setBatchSize, setBiases, setCudaHandles, setDeltas, setL1Regularization, setL2Regularization, setLearningRate, setMode, setMomentum, setNetworkType, setNextlayer, setOptimizerType, setOutputs, setPrevDeltaWeights, setPrevLayer, setThreadPool, setTrainable, setWeights, setWeights

Methods inherited from class Object

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

Constructor Details

MaxPoolingLayer

public MaxPoolingLayer(int filterWidth,
 int filterHeight,
 int stride)

Creates a new max pooling layer with specified filter dimensions and stride.

Parameters:

filterWidth - width of the filter square

filterHeight - height of the filter square

stride - filter step

MaxPoolingLayer

public MaxPoolingLayer(Filter filter)

Method Details

init

public final void init()

Description copied from class: AbstractLayer

This method should implement layer initialization in subclasses, when a layer is added to the network (create weights, outputs, deltas, randomization etc.).

Specified by:

init in class AbstractLayer<TensorBase, TensorBase, TensorBase>

initTransientFields

public void initTransientFields()

Overrides:

initTransientFields in class AbstractLayer<TensorBase, TensorBase, TensorBase>

forward

public void forward()

Max pooling forward pass outputs the max value for each filter position.

Specified by:

forward in interface Forward

Specified by:

forward in interface Layer<TensorBase>

Specified by:

forward in class AbstractLayer<TensorBase, TensorBase, TensorBase>

backward

public void backward()

backward pass for a max(x, y) operation has a simple interpretation as only routing the gradient to the input that had the highest value in the forward pass. Hence, during the forward pass of a pooling layer it is common to keep track of the index of the max activation (sometimes also called the switches) so that gradient routing is efficient during backpropagation.

Specified by:

backward in interface Backward

Specified by:

backward in interface Layer<TensorBase>

Specified by:

backward in class AbstractLayer<TensorBase, TensorBase, TensorBase>

applyWeightChanges

public void applyWeightChanges()

Does nothing for pooling layer since it does not have weights It just propagates deltas from next layer to previous through connections that had max activation in forward pass

Specified by:

applyWeightChanges in class AbstractLayer<TensorBase, TensorBase, TensorBase>

getFilterWidth

public int getFilterWidth()

getFilterHeight

public int getFilterHeight()

getStride

public int getStride()

getL1WeightSum

public float getL1WeightSum()

Overrides:

getL1WeightSum in class AbstractLayer<TensorBase, TensorBase, TensorBase>

getL2WeightSum

public float getL2WeightSum()

Overrides:

getL2WeightSum in class AbstractLayer<TensorBase, TensorBase, TensorBase>

toString

public String toString()

Overrides:

toString in class Object