org.deeplearning4j.nn.conf.layers

Class ConvolutionLayer.Builder

java.lang.Object

org.deeplearning4j.nn.conf.layers.Layer.Builder<T>

org.deeplearning4j.nn.conf.layers.FeedForwardLayer.Builder<T>

org.deeplearning4j.nn.conf.layers.ConvolutionLayer.BaseConvBuilder<ConvolutionLayer.Builder>

org.deeplearning4j.nn.conf.layers.ConvolutionLayer.Builder

Enclosing class:

ConvolutionLayer

public static class ConvolutionLayer.Builder
extends ConvolutionLayer.BaseConvBuilder<ConvolutionLayer.Builder>

Field Summary

Fields inherited from class org.deeplearning4j.nn.conf.layers.ConvolutionLayer.BaseConvBuilder

convolutionMode, cudnnAlgoMode, kernelSize, padding, stride

Fields inherited from class org.deeplearning4j.nn.conf.layers.FeedForwardLayer.Builder

nIn, nOut

Fields inherited from class org.deeplearning4j.nn.conf.layers.Layer.Builder

activationFn, adamMeanDecay, adamVarDecay, biasInit, biasLearningRate, dist, dropOut, epsilon, gradientNormalization, gradientNormalizationThreshold, l1, l1Bias, l2, l2Bias, layerName, learningRate, learningRatePolicy, learningRateSchedule, momentum, momentumAfter, rho, rmsDecay, updater, weightInit

Constructor Summary

Constructors

Constructor and Description
Builder()
Builder(int... kernelSize)
Builder(int[] kernelSize, int[] stride)
Builder(int[] kernelSize, int[] stride, int[] padding)

Method Summary

Modifier and Type	Method and Description
ConvolutionLayer.Builder	activation(org.nd4j.linalg.activations.Activation activation)
ConvolutionLayer.Builder	activation(org.nd4j.linalg.activations.IActivation activationFunction)
ConvolutionLayer.Builder	adamMeanDecay(double adamMeanDecay) Mean decay rate for Adam updater.
ConvolutionLayer.Builder	adamVarDecay(double adamVarDecay) Variance decay rate for Adam updater.
ConvolutionLayer.Builder	biasInit(double biasInit)
ConvolutionLayer.Builder	biasLearningRate(double biasLearningRate) Bias learning rate.
ConvolutionLayer	build()
ConvolutionLayer.Builder	convolutionMode(ConvolutionMode convolutionMode) Set the convolution mode for the Convolution layer.
ConvolutionLayer.Builder	cudnnAlgoMode(ConvolutionLayer.AlgoMode cudnnAlgoMode) Defaults to "PREFER_FASTEST", but "NO_WORKSPACE" uses less memory.
ConvolutionLayer.Builder	dist(Distribution dist) Distribution to sample initial weights from.
ConvolutionLayer.Builder	dropOut(double dropOut) Dropout.
ConvolutionLayer.Builder	epsilon(double epsilon) Epsilon value for updaters: Adagrad and Adadelta.
ConvolutionLayer.Builder	gradientNormalization(GradientNormalization gradientNormalization) Gradient normalization strategy.
ConvolutionLayer.Builder	gradientNormalizationThreshold(double threshold) Threshold for gradient normalization, only used for GradientNormalization.ClipL2PerLayer, GradientNormalization.ClipL2PerParamType, and GradientNormalization.ClipElementWiseAbsoluteValue Not used otherwise. L2 threshold for first two types of clipping, or absolute value threshold for last type of clipping.
ConvolutionLayer.Builder	kernelSize(int... kernelSize) Size of the convolution rows/columns
ConvolutionLayer.Builder	l1(double l1) L1 regularization coefficient (weights only).
ConvolutionLayer.Builder	l1Bias(double l1Bias) L1 regularization coefficient for the bias.
ConvolutionLayer.Builder	l2(double l2) L2 regularization coefficient (weights only).
ConvolutionLayer.Builder	l2Bias(double l2Bias) L2 regularization coefficient for the bias.
ConvolutionLayer.Builder	learningRate(double learningRate) Learning rate.
ConvolutionLayer.Builder	learningRateDecayPolicy(LearningRatePolicy policy) Learning rate decay policy.
ConvolutionLayer.Builder	learningRateSchedule(java.util.Map<java.lang.Integer,java.lang.Double> learningRateSchedule) Learning rate schedule.
ConvolutionLayer.Builder	momentum(double momentum) Momentum rate.
ConvolutionLayer.Builder	momentumAfter(java.util.Map<java.lang.Integer,java.lang.Double> momentumAfter) Momentum schedule.
ConvolutionLayer.Builder	name(java.lang.String layerName) Layer name assigns layer string name.
ConvolutionLayer.Builder	nIn(int nIn)
ConvolutionLayer.Builder	nOut(int nOut)
ConvolutionLayer.Builder	padding(int... padding)
ConvolutionLayer.Builder	rho(double rho) Ada delta coefficient, rho.
ConvolutionLayer.Builder	rmsDecay(double rmsDecay) Decay rate for RMSProp.
ConvolutionLayer.Builder	stride(int... stride)
ConvolutionLayer.Builder	updater(Updater updater) Gradient updater.
ConvolutionLayer.Builder	weightInit(WeightInit weightInit) Weight initialization scheme.

Methods inherited from class org.deeplearning4j.nn.conf.layers.Layer.Builder

activation

Methods inherited from class java.lang.Object

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

Constructor Detail

Builder

public Builder(int[] kernelSize,
               int[] stride,
               int[] padding)

Builder

public Builder(int[] kernelSize,
               int[] stride)

Builder

public Builder(int... kernelSize)

Builder

public Builder()

Method Detail

convolutionMode

public ConvolutionLayer.Builder convolutionMode(ConvolutionMode convolutionMode)

Set the convolution mode for the Convolution layer. See ConvolutionMode for more details

Overrides:

convolutionMode in class ConvolutionLayer.BaseConvBuilder<ConvolutionLayer.Builder>

Parameters:

convolutionMode - Convolution mode for layer

nIn

public ConvolutionLayer.Builder nIn(int nIn)

Overrides:

nIn in class FeedForwardLayer.Builder<ConvolutionLayer.Builder>

nOut

public ConvolutionLayer.Builder nOut(int nOut)

Overrides:

nOut in class FeedForwardLayer.Builder<ConvolutionLayer.Builder>

cudnnAlgoMode

public ConvolutionLayer.Builder cudnnAlgoMode(ConvolutionLayer.AlgoMode cudnnAlgoMode)

Defaults to "PREFER_FASTEST", but "NO_WORKSPACE" uses less memory.

Overrides:

cudnnAlgoMode in class ConvolutionLayer.BaseConvBuilder<ConvolutionLayer.Builder>

Parameters:

cudnnAlgoMode -

name

public ConvolutionLayer.Builder name(java.lang.String layerName)

Layer name assigns layer string name. Allows easier differentiation between layers.

Overrides:

name in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

layerName -

activation

public ConvolutionLayer.Builder activation(org.nd4j.linalg.activations.IActivation activationFunction)

Overrides:

activation in class Layer.Builder<ConvolutionLayer.Builder>

activation

public ConvolutionLayer.Builder activation(org.nd4j.linalg.activations.Activation activation)

Overrides:

activation in class Layer.Builder<ConvolutionLayer.Builder>

weightInit

public ConvolutionLayer.Builder weightInit(WeightInit weightInit)

Weight initialization scheme.

Overrides:

weightInit in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

weightInit -

See Also:

WeightInit

biasInit

public ConvolutionLayer.Builder biasInit(double biasInit)

Overrides:

biasInit in class Layer.Builder<ConvolutionLayer.Builder>

dist

public ConvolutionLayer.Builder dist(Distribution dist)

Distribution to sample initial weights from. Used in conjunction with .weightInit(WeightInit.DISTRIBUTION).

Overrides:

dist in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

dist -

learningRate

public ConvolutionLayer.Builder learningRate(double learningRate)

Learning rate. Defaults to 1e-1

Overrides:

learningRate in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

learningRate -

biasLearningRate

public ConvolutionLayer.Builder biasLearningRate(double biasLearningRate)

Bias learning rate. Set this to apply a different learning rate to the bias

Overrides:

biasLearningRate in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

biasLearningRate -

learningRateSchedule

public ConvolutionLayer.Builder learningRateSchedule(java.util.Map<java.lang.Integer,java.lang.Double> learningRateSchedule)

Learning rate schedule. Map of the iteration to the learning rate to apply at that iteration.

Overrides:

learningRateSchedule in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

learningRateSchedule -

l1

public ConvolutionLayer.Builder l1(double l1)

L1 regularization coefficient (weights only). Use l1Bias(double) to configure the l1 regularization coefficient for the bias.

Overrides:

l1 in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

l1 -

l2

public ConvolutionLayer.Builder l2(double l2)

L2 regularization coefficient (weights only). Use l2Bias(double) to configure the l2 regularization coefficient for the bias.

Overrides:

l2 in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

l2 -

l1Bias

public ConvolutionLayer.Builder l1Bias(double l1Bias)

L1 regularization coefficient for the bias. Default: 0. See also l1(double)

Overrides:

l1Bias in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

l1Bias -

l2Bias

public ConvolutionLayer.Builder l2Bias(double l2Bias)

L2 regularization coefficient for the bias. Default: 0. See also l2(double)

Overrides:

l2Bias in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

l2Bias -

dropOut

public ConvolutionLayer.Builder dropOut(double dropOut)

Dropout. Value is probability of retaining an activation - thus 1.0 is equivalent to no dropout. Note that 0.0 (the default) disables dropout.

Overrides:

dropOut in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

dropOut -

momentum

public ConvolutionLayer.Builder momentum(double momentum)

Momentum rate.

Overrides:

momentum in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

momentum -

momentumAfter

public ConvolutionLayer.Builder momentumAfter(java.util.Map<java.lang.Integer,java.lang.Double> momentumAfter)

Momentum schedule. Map of the iteration to the momentum rate to apply at that iteration.

Overrides:

momentumAfter in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

momentumAfter -

updater

public ConvolutionLayer.Builder updater(Updater updater)

Gradient updater. For example, SGD for standard stochastic gradient descent, NESTEROV for Nesterov momentum, RSMPROP for RMSProp, etc.

Overrides:

updater in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

updater -

See Also:

Updater

rho

public ConvolutionLayer.Builder rho(double rho)

Ada delta coefficient, rho. Only applies if using .updater(Updater.ADADELTA)

Overrides:

rho in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

rho -

rmsDecay

public ConvolutionLayer.Builder rmsDecay(double rmsDecay)

Decay rate for RMSProp. Only applies if using .updater(Updater.RMSPROP)

Overrides:

rmsDecay in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

rmsDecay -

epsilon

public ConvolutionLayer.Builder epsilon(double epsilon)

Epsilon value for updaters: Adagrad and Adadelta. Only used if using Updater.ADAGRAD or Updater.ADADELTA

Overrides:

epsilon in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

epsilon - Epsilon value to use for adagrad and adadelta

adamMeanDecay

public ConvolutionLayer.Builder adamMeanDecay(double adamMeanDecay)

Mean decay rate for Adam updater. Only applies if using .updater(Updater.ADAM)

Overrides:

adamMeanDecay in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

adamMeanDecay -

adamVarDecay

public ConvolutionLayer.Builder adamVarDecay(double adamVarDecay)

Variance decay rate for Adam updater. Only applies if using .updater(Updater.ADAM)

Overrides:

adamVarDecay in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

adamVarDecay -

gradientNormalization

public ConvolutionLayer.Builder gradientNormalization(GradientNormalization gradientNormalization)

Gradient normalization strategy. Used to specify gradient renormalization, gradient clipping etc.

Overrides:

gradientNormalization in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

gradientNormalization - Type of normalization to use. Defaults to None.

See Also:

GradientNormalization

gradientNormalizationThreshold

public ConvolutionLayer.Builder gradientNormalizationThreshold(double threshold)

Threshold for gradient normalization, only used for GradientNormalization.ClipL2PerLayer, GradientNormalization.ClipL2PerParamType, and GradientNormalization.ClipElementWiseAbsoluteValue
Not used otherwise.
L2 threshold for first two types of clipping, or absolute value threshold for last type of clipping.

Overrides:

gradientNormalizationThreshold in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

threshold -

learningRateDecayPolicy

public ConvolutionLayer.Builder learningRateDecayPolicy(LearningRatePolicy policy)

Learning rate decay policy. Used to adapt learning rate based on policy.

Overrides:

learningRateDecayPolicy in class Layer.Builder<ConvolutionLayer.Builder>

Parameters:

policy - Type of policy to use. Defaults to None.

See Also:

GradientNormalization

kernelSize

public ConvolutionLayer.Builder kernelSize(int... kernelSize)

Size of the convolution rows/columns

Parameters:

kernelSize - the height and width of the kernel

Returns:

stride

public ConvolutionLayer.Builder stride(int... stride)

padding

public ConvolutionLayer.Builder padding(int... padding)

build

public ConvolutionLayer build()

Specified by:

build in class Layer.Builder<ConvolutionLayer.Builder>