org.deeplearning4j.nn.layers.training

Class CenterLossOutputLayer

java.lang.Object

org.deeplearning4j.nn.layers.AbstractLayer<LayerConfT>

org.deeplearning4j.nn.layers.BaseLayer<LayerConfT>

org.deeplearning4j.nn.layers.BaseOutputLayer<CenterLossOutputLayer>

org.deeplearning4j.nn.layers.training.CenterLossOutputLayer

All Implemented Interfaces:

Serializable, Cloneable, Classifier, Layer, IOutputLayer, Model

public class CenterLossOutputLayer
extends BaseOutputLayer<CenterLossOutputLayer>

Center loss is similar to triplet loss except that it enforces intraclass consistency and doesn't require feed forward of multiple examples. Center loss typically converges faster for training ImageNet-based convolutional networks. "If example x is in class Y, ensure that embedding(x) is close to average(embedding(y)) for all examples y in Y"

Author:

Justin Long (@crockpotveggies)

See Also:

Serialized Form

Nested Class Summary

Nested classes/interfaces inherited from interface org.deeplearning4j.nn.api.Layer

Layer.TrainingMode, Layer.Type

Field Summary

Fields inherited from class org.deeplearning4j.nn.layers.BaseOutputLayer

inputMaskArray, inputMaskArrayState, labels

Fields inherited from class org.deeplearning4j.nn.layers.BaseLayer

gradient, gradientsFlattened, gradientViews, optimizer, params, paramsFlattened, score, weightNoiseParams

Fields inherited from class org.deeplearning4j.nn.layers.AbstractLayer

cacheMode, conf, dropoutApplied, dropoutMask, epochCount, index, input, iterationCount, iterationListeners, maskArray, maskState, preOutput

Constructor Summary

Constructors

Constructor and Description
CenterLossOutputLayer(NeuralNetConfiguration conf)
CenterLossOutputLayer(NeuralNetConfiguration conf, org.nd4j.linalg.api.ndarray.INDArray input)

Method Summary

Modifier and Type	Method and Description
org.nd4j.linalg.primitives.Pair<Gradient,org.nd4j.linalg.api.ndarray.INDArray>	backpropGradient(org.nd4j.linalg.api.ndarray.INDArray epsilon) Calculate the gradient relative to the error in the next layer
void	computeGradientAndScore() Update the score
double	computeScore(double fullNetworkL1, double fullNetworkL2, boolean training) Compute score after labels and input have been set.
org.nd4j.linalg.api.ndarray.INDArray	computeScoreForExamples(double fullNetworkL1, double fullNetworkL2) Compute the score for each example individually, after labels and input have been set.
Gradient	gradient() Gets the gradient from one training iteration
org.nd4j.linalg.primitives.Pair<Gradient,Double>	gradientAndScore() Get the gradient and score
protected void	setScoreWithZ(org.nd4j.linalg.api.ndarray.INDArray z)

Methods inherited from class org.deeplearning4j.nn.layers.BaseOutputLayer

activate, activate, activate, applyMask, clear, f1Score, f1Score, fit, fit, fit, fit, fit, getLabels, getLabels2d, hasBias, isPretrainLayer, iterate, labelProbabilities, migrateInput, numLabels, output, output, output, predict, predict, preOutput2d, setLabels

Methods inherited from class org.deeplearning4j.nn.layers.BaseLayer

accumulateScore, activate, activate, activate, calcL1, calcL2, clearNoiseWeightParams, clone, fit, getGradientsViewArray, getOptimizer, getParam, getParamWithNoise, initParams, layerConf, numParams, params, paramTable, paramTable, preOutput, preOutput, score, setBackpropGradientsViewArray, setParam, setParams, setParams, setParamsViewArray, setParamTable, toString, transpose, update, update

Methods inherited from class org.deeplearning4j.nn.layers.AbstractLayer

addListeners, applyConstraints, applyDropOutIfNecessary, batchSize, conf, feedForwardMaskArray, getIndex, getInput, getInputMiniBatchSize, getListeners, getMaskArray, init, input, layerId, numParams, preOutput, preOutput, setCacheMode, setConf, setIndex, setInput, setInputMiniBatchSize, setListeners, setListeners, setMaskArray, type, validateInput

Methods inherited from class java.lang.Object

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

Methods inherited from interface org.deeplearning4j.nn.api.Layer

activate, activate, activate, calcL1, calcL2, clearNoiseWeightParams, clone, feedForwardMaskArray, getEpochCount, getIndex, getInputMiniBatchSize, getIterationCount, getListeners, getMaskArray, preOutput, preOutput, preOutput, setCacheMode, setEpochCount, setIndex, setInput, setInputMiniBatchSize, setIterationCount, setListeners, setListeners, setMaskArray, transpose, type

Methods inherited from interface org.deeplearning4j.nn.api.Model

accumulateScore, addListeners, applyConstraints, batchSize, conf, fit, getGradientsViewArray, getOptimizer, getParam, init, initParams, input, numParams, numParams, params, paramTable, paramTable, score, setBackpropGradientsViewArray, setConf, setParam, setParams, setParamsViewArray, setParamTable, update, update, validateInput

Constructor Detail

CenterLossOutputLayer

public CenterLossOutputLayer(NeuralNetConfiguration conf)

CenterLossOutputLayer

public CenterLossOutputLayer(NeuralNetConfiguration conf,
                             org.nd4j.linalg.api.ndarray.INDArray input)

Method Detail

computeScore

public double computeScore(double fullNetworkL1,
                           double fullNetworkL2,
                           boolean training)

Compute score after labels and input have been set.

Specified by:

computeScore in interface IOutputLayer

Overrides:

computeScore in class BaseOutputLayer<CenterLossOutputLayer>

Parameters:

fullNetworkL1 - L1 regularization term for the entire network

fullNetworkL2 - L2 regularization term for the entire network

training - whether score should be calculated at train or test time (this affects things like application of dropout, etc)

Returns:

score (loss function)

computeScoreForExamples

public org.nd4j.linalg.api.ndarray.INDArray computeScoreForExamples(double fullNetworkL1,
                                                                    double fullNetworkL2)

Compute the score for each example individually, after labels and input have been set.

Specified by:

computeScoreForExamples in interface IOutputLayer

Overrides:

computeScoreForExamples in class BaseOutputLayer<CenterLossOutputLayer>

Parameters:

fullNetworkL1 - L1 regularization term for the entire network (or, 0.0 to not include regularization)

fullNetworkL2 - L2 regularization term for the entire network (or, 0.0 to not include regularization)

Returns:

A column INDArray of shape [numExamples,1], where entry i is the score of the ith example

computeGradientAndScore

public void computeGradientAndScore()

Description copied from interface: Model

Update the score

Specified by:

computeGradientAndScore in interface Model

Overrides:

computeGradientAndScore in class BaseOutputLayer<CenterLossOutputLayer>

setScoreWithZ

protected void setScoreWithZ(org.nd4j.linalg.api.ndarray.INDArray z)

Overrides:

setScoreWithZ in class BaseOutputLayer<CenterLossOutputLayer>

gradientAndScore

public org.nd4j.linalg.primitives.Pair<Gradient,Double> gradientAndScore()

Description copied from interface: Model

Get the gradient and score

Specified by:

gradientAndScore in interface Model

Overrides:

gradientAndScore in class BaseOutputLayer<CenterLossOutputLayer>

Returns:

the gradient and score

backpropGradient

public org.nd4j.linalg.primitives.Pair<Gradient,org.nd4j.linalg.api.ndarray.INDArray> backpropGradient(org.nd4j.linalg.api.ndarray.INDArray epsilon)

Description copied from interface: Layer

Calculate the gradient relative to the error in the next layer

Specified by:

backpropGradient in interface Layer

Overrides:

backpropGradient in class BaseOutputLayer<CenterLossOutputLayer>

Parameters:

epsilon - w^(L+1)*delta^(L+1). Or, equiv: dC/da, i.e., (dC/dz)*(dz/da) = dC/da, where C is cost function a=sigma(z) is activation.

Returns:

Pair where Gradient is gradient for this layer, INDArray is epsilon needed by next layer, but before element-wise multiply by sigmaPrime(z). So for standard feed-forward layer, if this layer is L, then return.getSecond() == (w^(L)*(delta^(L))^T)^T

gradient

public Gradient gradient()

Gets the gradient from one training iteration

Specified by:

gradient in interface Model

Overrides:

gradient in class BaseOutputLayer<CenterLossOutputLayer>

Returns:

the gradient (bias and weight matrix)