Using @SuperBuilder for LayerConfigurations

Signed-off-by: brian <brian@brutex.de>
master
Brian Rosenberger 2023-04-25 15:42:24 +02:00
parent 391a1ad397
commit 8f524827e4
40 changed files with 427 additions and 404 deletions

View File

@ -171,7 +171,7 @@ public class App {
LayerConfiguration[] disLayers = Arrays.stream(disLayers())
.map((layer) -> {
if (layer instanceof DenseLayer || layer instanceof OutputLayer) {
return new FrozenLayerWithBackprop(layer);
return FrozenLayerWithBackprop.builder(layer);
} else {
return layer;
}

View File

@ -162,20 +162,20 @@ public class UtilLayerGradientChecks extends BaseDL4JTest {
}
NeuralNetConfiguration conf = NeuralNetConfiguration.builder()
NeuralNetConfiguration conf =
NeuralNetConfiguration.builder()
.updater(new NoOp())
.activation(Activation.TANH)
.dataType(DataType.DOUBLE)
.dist(new NormalDistribution(0,2))
.dist(new NormalDistribution(0, 2))
.list()
.layer(l1)
.layer(new MaskLayer())
.layer(MaskLayer.builder().build())
.layer(l2)
.layer(l3)
.inputType(it)
.build();
MultiLayerNetwork net = new MultiLayerNetwork(conf);
net.init();
@ -203,11 +203,11 @@ public class UtilLayerGradientChecks extends BaseDL4JTest {
.list()
.layer(DenseLayer.builder().nIn(10).nOut(10)
.activation(Activation.TANH).weightInit(WeightInit.XAVIER).build())
.layer(new FrozenLayerWithBackprop(DenseLayer.builder().nIn(10).nOut(10)
.activation(Activation.TANH).weightInit(WeightInit.XAVIER).build()))
.layer(new FrozenLayerWithBackprop(
.layer(FrozenLayerWithBackprop.builder().underlying(DenseLayer.builder().nIn(10).nOut(10)
.activation(Activation.TANH).weightInit(WeightInit.XAVIER).build()).build())
.layer(FrozenLayerWithBackprop.builder().underlying(
DenseLayer.builder().nIn(10).nOut(10).activation(Activation.TANH)
.weightInit(WeightInit.XAVIER).build()))
.weightInit(WeightInit.XAVIER).build()).build())
.layer(OutputLayer.builder().lossFunction(LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nIn(10).nOut(10).build())
.build();

View File

@ -40,15 +40,12 @@ import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.learning.config.IUpdater;
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
public class ActivationLayer extends NoParamLayer {
{
setType(LayerType.ACT);
}
public static ActivationLayerBuilder<?, ?> builder(Activation activation) {
return innerBuilder().activation(activation);

View File

@ -49,6 +49,8 @@ import org.nd4j.linalg.learning.regularization.WeightDecay;
@SuperBuilder
public abstract class BaseLayerConfiguration extends LayerConfiguration
implements ITraininableLayerConfiguration, Serializable, Cloneable {
/**
* Set constraints to be applied to all layers. Default: no constraints.<br>
* Constraints can be used to enforce certain conditions (non-negativity of parameters, max-norm
@ -84,9 +86,9 @@ public abstract class BaseLayerConfiguration extends LayerConfiguration
@Getter @Setter @Builder.Default
protected double gainInit = 0.0;
/** Regularization for the parameters (excluding biases). */
@Builder.Default @Getter protected List<Regularization> regularization = new ArrayList<>();
@Builder.Default @Getter @Setter protected List<Regularization> regularization = new ArrayList<>();
/** Regularization for the bias parameters only */
@Builder.Default @Getter
@Builder.Default @Getter @Setter
protected List<Regularization> regularizationBias = new ArrayList<>();
/**
* Gradient updater. For example, {@link org.nd4j.linalg.learning.config.Adam} or {@link
@ -210,6 +212,7 @@ public abstract class BaseLayerConfiguration extends LayerConfiguration
C extends BaseLayerConfiguration, B extends BaseLayerConfigurationBuilder<C, B>>
extends LayerConfigurationBuilder<C, B> {
/**
* Set weight initialization scheme to random sampling via the specified distribution.
* Equivalent to: {@code .weightInit(new WeightInitDistribution(distribution))}

View File

@ -29,8 +29,7 @@ import org.nd4j.linalg.lossfunctions.ILossFunction;
import org.nd4j.linalg.lossfunctions.LossFunctions;
import org.nd4j.linalg.lossfunctions.impl.LossMCXENT;
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(builderMethodName = "innerBuilder")
@ -39,19 +38,16 @@ public abstract class BaseOutputLayer extends FeedForwardLayer {
/**
* Loss function for the output layer
*/
@lombok.Builder.Default
@lombok.Builder.Default @Getter @Setter
protected ILossFunction lossFunction = new LossMCXENT();
/**
* If true (default): include bias parameters in the model. False: no bias.
*
*/
@lombok.Builder.Default
@lombok.Builder.Default @Getter @Setter
protected boolean hasBias = true;
public boolean hasBias() {
return hasBias;
}
@Override
public LayerMemoryReport getMemoryReport(InputType inputType) {

View File

@ -31,11 +31,11 @@ import org.nd4j.linalg.lossfunctions.LossFunctions;
@JsonIgnoreProperties("pretrain")
@SuperBuilder
public abstract class BasePretrainNetwork extends FeedForwardLayer {
@Builder.Default
@Builder.Default @Getter
protected LossFunctions.LossFunction lossFunction =
LossFunctions.LossFunction.RECONSTRUCTION_CROSSENTROPY;
@Builder.Default protected double visibleBiasInit = 0.0;
@Builder.Default @Getter protected double visibleBiasInit = 0.0;
@Override
public boolean isPretrainParam(String paramName) {

View File

@ -31,8 +31,6 @@ import org.deeplearning4j.nn.conf.inputs.InputType;
* @author Max Pumperla
*/
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder()
@ -43,7 +41,7 @@ public abstract class BaseUpsamplingLayer extends NoParamLayer {
* dimensions (e.g. 2 for Upsampling2D etc.)
*
*/
@Builder.Default
@Builder.Default @Getter
protected int[] size = new int[] {1};
@Override
@ -60,8 +58,4 @@ public abstract class BaseUpsamplingLayer extends NoParamLayer {
}
return InputTypeUtil.getPreProcessorForInputTypeCnnLayers(inputType, getName());
}
}

View File

@ -42,7 +42,6 @@ import java.util.Collection;
import java.util.Map;
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")

View File

@ -37,15 +37,15 @@ import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.lossfunctions.ILossFunction;
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder
public class Cnn3DLossLayer extends FeedForwardLayer {
@Getter @Setter
protected ILossFunction lossFunction;
/** Format of the input/output data. See {@link Convolution3D.DataFormat} for details */
@Getter @Setter
protected Convolution3D.DataFormat dataFormat;
@Override

View File

@ -24,10 +24,11 @@ import lombok.Data;
import lombok.EqualsAndHashCode;
import lombok.NoArgsConstructor;
import lombok.ToString;
import lombok.experimental.SuperBuilder;
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder
public class Convolution1D extends Convolution1DLayer {
}

View File

@ -45,7 +45,6 @@ import org.nd4j.linalg.api.ndarray.INDArray;
* wide.
*/
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@ -142,7 +141,7 @@ public class Convolution1DLayer extends ConvolutionLayer {
} else {
outLength =
Convolution1DUtils.getOutputSize(
inputTsLength, kernelSize[0], stride[0], padding[0], convolutionMode, dilation[0]);
inputTsLength, kernelSize[0], stride[0], padding[0], getConvolutionMode(), dilation[0]);
}
return InputType.recurrent(nOut, outLength, rnnDataFormat);

View File

@ -24,10 +24,12 @@ import lombok.Data;
import lombok.EqualsAndHashCode;
import lombok.NoArgsConstructor;
import lombok.ToString;
import lombok.experimental.SuperBuilder;
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder
public class Convolution2D extends ConvolutionLayer {
}

View File

@ -38,7 +38,6 @@ import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(builderMethodName = "innerBuilder", buildMethodName = "initBuild")
@ -118,7 +117,7 @@ public class Convolution3D extends ConvolutionLayer {
* kernel size
*/
public boolean hasBias() {
return hasBias;
return isHasBias();
}
@Override

View File

@ -46,6 +46,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
* to be used in the net or in other words the channels The builder specifies the filter/kernel
* size, the stride and padding The pooling layer takes the kernel size
*/
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@ -55,14 +56,14 @@ public class ConvolutionLayer extends FeedForwardLayer {
*
* @param kernelSize the height and width of the kernel
*/
public @Builder.Default int[] kernelSize = new int[] {5, 5}; // Square filter
private @Builder.Default @Getter @Setter int[] kernelSize = new int[] {5, 5}; // Square filter
/** If true (default): include bias parameters in the model. False: no bias. */
@Builder.Default protected boolean hasBias = true;
@Builder.Default @Getter @Setter private boolean hasBias = true;
/**
* Set the convolution mode for the Convolution layer. See {@link ConvolutionMode} for more
* details Default is {@link ConvolutionMode}.Truncate.
*/
@Builder.Default protected ConvolutionMode convolutionMode = ConvolutionMode.Truncate;
@Builder.Default @Getter @Setter private ConvolutionMode convolutionMode = ConvolutionMode.Truncate;
/**
* Set the data format for the CNN activations - NCHW (channels first) or NHWC (channels last).
@ -72,7 +73,7 @@ public class ConvolutionLayer extends FeedForwardLayer {
* @param format Format for activations (in and out)
*/
@Builder.Default
protected CNN2DFormat convFormat =
private CNN2DFormat convFormat =
CNN2DFormat.NCHW; // default value for legacy serialization reasons
/**
@ -85,25 +86,25 @@ public class ConvolutionLayer extends FeedForwardLayer {
* http://deeplearning.net/software/theano/tutorial/conv_arithmetic.html#dilated-convolutions</a>
* <br>
*/
protected @Builder.Default int[] dilation = new int[] {1, 1};
private @Builder.Default int[] dilation = new int[] {1, 1};
/** Default is 2. Down-sample by a factor of 2 */
protected @Builder.Default int[] stride = new int[] {1, 1};
private @Builder.Default int[] stride = new int[] {1, 1};
protected @Builder.Default int[] padding = new int[] {0, 0};
private @Builder.Default int[] padding = new int[] {0, 0};
/**
* When using CuDNN and an error is encountered, should fallback to the non-CuDNN implementatation
* be allowed? If set to false, an exception in CuDNN will be propagated back to the user. If
* false, the built-in (non-CuDNN) implementation for ConvolutionLayer will be used
*/
@Builder.Default protected boolean cudnnAllowFallback = true;
@Builder.Default private boolean cudnnAllowFallback = true;
/** Defaults to "PREFER_FASTEST", but "NO_WORKSPACE" uses less memory. */
@Builder.Default protected AlgoMode cudnnAlgoMode = AlgoMode.PREFER_FASTEST;
@Builder.Default private AlgoMode cudnnAlgoMode = AlgoMode.PREFER_FASTEST;
protected FwdAlgo cudnnFwdAlgo;
protected BwdFilterAlgo cudnnBwdFilterAlgo;
protected BwdDataAlgo cudnnBwdDataAlgo;
@Builder.Default protected int convolutionDim = 2; // 2D convolution by default
private FwdAlgo cudnnFwdAlgo;
private BwdFilterAlgo cudnnBwdFilterAlgo;
private BwdDataAlgo cudnnBwdDataAlgo;
@Builder.Default private int convolutionDim = 2; // 2D convolution by default
/** Causal convolution - allowed for 1D only */
@Builder.Default private boolean allowCausal = false;

View File

@ -44,7 +44,6 @@ import java.util.Map;
* The pooling layer takes the kernel size
*/
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuild")
@ -88,20 +87,20 @@ private CNN2DFormat format = CNN2DFormat.NCHW;
}
}
public boolean hasBias() {
return hasBias;
return isHasBias();
}
@Override
public Deconvolution2D clone() {
Deconvolution2D clone = (Deconvolution2D) super.clone();
if (clone.kernelSize != null) {
clone.kernelSize = clone.kernelSize.clone();
if (clone.getKernelSize() != null) {
clone.setKernelSize( clone.getKernelSize().clone());
}
if (clone.stride != null) {
clone.stride = clone.stride.clone();
if (clone.getStride() != null) {
clone.setStride( clone.getStride().clone());
}
if (clone.padding != null) {
clone.padding = clone.padding.clone();
if (clone.getPadding() != null) {
clone.setPadding( clone.getPadding().clone());
}
return clone;
}
@ -138,7 +137,7 @@ private CNN2DFormat format = CNN2DFormat.NCHW;
+ "\"): Expected CNN input, got " + inputType);
}
return InputTypeUtil.getOutputTypeDeconvLayer(inputType, kernelSize, stride, padding, dilation, convolutionMode,
return InputTypeUtil.getOutputTypeDeconvLayer(inputType, getKernelSize(), getStride(), getPadding(), getDilation(), getConvolutionMode(),
nOut, layerIndex, getName(), Deconvolution2DLayer.class);
}

View File

@ -42,7 +42,6 @@ import org.nd4j.linalg.api.ndarray.INDArray;
* filter/kernel size, the stride and padding The pooling layer takes the kernel size
*/
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@ -63,20 +62,20 @@ public class Deconvolution3D extends ConvolutionLayer {
}
public boolean hasBias() {
return hasBias;
return isHasBias();
}
@Override
public Deconvolution3D clone() {
Deconvolution3D clone = (Deconvolution3D) super.clone();
if (clone.kernelSize != null) {
clone.kernelSize = clone.kernelSize.clone();
if (clone.getKernelSize() != null) {
clone.setKernelSize( clone.getKernelSize().clone());
}
if (clone.stride != null) {
clone.stride = clone.stride.clone();
if (clone.getStride() != null) {
clone.setStride( clone.getStride().clone());
}
if (clone.padding != null) {
clone.padding = clone.padding.clone();
if (clone.getPadding() != null) {
clone.setPadding( clone.getPadding().clone());
}
return clone;
}
@ -147,11 +146,11 @@ public class Deconvolution3D extends ConvolutionLayer {
return InputTypeUtil.getOutputTypeDeconv3dLayer(
inputType,
kernelSize,
stride,
padding,
dilation,
convolutionMode,
getKernelSize(),
getStride(),
getPadding(),
getDilation(),
getConvolutionMode(),
dataFormat,
nOut,
layerIndex,

View File

@ -38,7 +38,7 @@ import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
/** Dense Layer Uses WeightInitXavier as default */
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(
@ -47,9 +47,9 @@ import org.nd4j.linalg.api.ndarray.INDArray;
public class DenseLayer extends FeedForwardLayer {
/** If true (default = false): enable layer normalization on this layer */
@lombok.Builder.Default @Accessors private boolean hasLayerNorm = false;
@lombok.Builder.Default private boolean hasLayerNorm = false;
@lombok.Builder.Default @Accessors private boolean hasBias = true;
@lombok.Builder.Default private boolean hasBias = true;
@Override
public Layer instantiate(

View File

@ -20,6 +20,7 @@
package org.deeplearning4j.nn.conf.layers;
import java.util.*;
import lombok.*;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.Layer;
@ -36,10 +37,7 @@ import org.nd4j.common.base.Preconditions;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import java.util.*;
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@ -47,12 +45,11 @@ public class DepthwiseConvolution2D extends ConvolutionLayer {
/**
* Set channels multiplier for depth-wise convolution
*
* @param depthMultiplier integer value, for each input map we get depthMultiplier outputs in channels-wise
* step.
* @param depthMultiplier integer value, for each input map we get depthMultiplier outputs in
* channels-wise step.
* @return Builder
*/
@Builder.Default
protected int depthMultiplier = 1;
@Builder.Default protected int depthMultiplier = 1;
/**
* Set the data format for the CNN activations - NCHW (channels first) or NHWC (channels last).
* See {@link CNN2DFormat} for more details.<br>
@ -67,50 +64,16 @@ public class DepthwiseConvolution2D extends ConvolutionLayer {
* Set the data format for the CNN activations - NCHW (channels first) or NHWC (channels last).
* See {@link CNN2DFormat} for more details.<br>
* Default: NCHW
*
* @param format Format for activations (in and out)
*/
@Builder.Default
protected CNN2DFormat cnn2DFormat = CNN2DFormat.NCHW;
public static abstract class DepthwiseConvolution2DBuilder<C extends DepthwiseConvolution2D, B extends DepthwiseConvolution2DBuilder<C, B>>
extends ConvolutionLayerBuilder<C, B> {
public C build() {
Preconditions.checkState(depthMultiplier$value > 0, "Depth multiplier must be > 0, got %s", depthMultiplier$value);
C l = this.initBuild();
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), l.getPadding());
ConvolutionUtils.validateCnnKernelStridePadding(l.getKernelSize(), l.getStride(), l.getPadding());
l.initializeConstraints();
return l;
}
@Override
public B kernelSize(int... kernelSize) {
super.kernelSize(ValidationUtils.validate2NonNegative(kernelSize, false, "kernelSize"));
return self();
}
@Override
public B stride(int... stride) {
super.stride(ValidationUtils.validate2NonNegative(stride, false, "stride"));
return self();
}
@Override
public B padding(int... padding) {
super.padding(ValidationUtils.validate2NonNegative(padding, false, "padding"));
return self();
}
@Override
public B dilation(int... dilation) {
super.dilation(ValidationUtils.validate2NonNegative(dilation, false, "dilation"));
return self();
}
}
@Builder.Default protected CNN2DFormat cnn2DFormat = CNN2DFormat.NCHW;
protected boolean allowCausal() {
//Causal convolution - allowed for 1D only
// Causal convolution - allowed for 1D only
return false;
}
@Override
public DepthwiseConvolution2D clone() {
DepthwiseConvolution2D clone = (DepthwiseConvolution2D) super.clone();
@ -118,11 +81,16 @@ public class DepthwiseConvolution2D extends ConvolutionLayer {
return clone;
}
@Override
public Layer instantiate(NeuralNetConfiguration conf, Collection<TrainingListener> trainingListeners,
int layerIndex, INDArray layerParamsView, boolean initializeParams, DataType networkDataType) {
LayerValidation.assertNInNOutSet("DepthwiseConvolution2D", getName(), layerIndex, getNIn(), getNOut());
public Layer instantiate(
NeuralNetConfiguration conf,
Collection<TrainingListener> trainingListeners,
int layerIndex,
INDArray layerParamsView,
boolean initializeParams,
DataType networkDataType) {
LayerValidation.assertNInNOutSet(
"DepthwiseConvolution2D", getName(), layerIndex, getNIn(), getNOut());
LayerConfiguration lconf = conf.getFlattenedLayerConfigurations().get(layerIndex);
runInheritance();
@ -146,24 +114,75 @@ public class DepthwiseConvolution2D extends ConvolutionLayer {
@Override
public InputType getOutputType(int layerIndex, InputType inputType) {
if (inputType == null || inputType.getType() != InputType.Type.CNN) {
throw new IllegalStateException("Invalid input for depth-wise convolution layer (layer name=\""
+ getName() + "\"): Expected CNN input, got " + inputType);
throw new IllegalStateException(
"Invalid input for depth-wise convolution layer (layer name=\""
+ getName()
+ "\"): Expected CNN input, got "
+ inputType);
}
return InputTypeUtil.getOutputTypeCnnLayers(inputType, kernelSize, stride, padding, dilation, convolutionMode,
nOut, layerIndex, getName(), dataFormat, DepthwiseConvolution2DLayer.class);
return InputTypeUtil.getOutputTypeCnnLayers(
inputType,
getKernelSize(),
getStride(),
getPadding(),
getDilation(),
getConvolutionMode(),
nOut,
layerIndex,
getName(),
dataFormat,
DepthwiseConvolution2DLayer.class);
}
@Override
public void setNIn(InputType inputType, boolean override) {
super.setNIn(inputType, override);
if(nOut == 0 || override){
if (nOut == 0 || override) {
nOut = this.nIn * this.depthMultiplier;
}
this.dataFormat = ((InputType.InputTypeConvolutional)inputType).getFormat();
this.dataFormat = ((InputType.InputTypeConvolutional) inputType).getFormat();
}
public abstract static class DepthwiseConvolution2DBuilder<
C extends DepthwiseConvolution2D, B extends DepthwiseConvolution2DBuilder<C, B>>
extends ConvolutionLayerBuilder<C, B> {
public C build() {
Preconditions.checkState(
depthMultiplier$value > 0,
"Depth multiplier must be > 0, got %s",
depthMultiplier$value);
C l = this.initBuild();
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), l.getPadding());
ConvolutionUtils.validateCnnKernelStridePadding(
l.getKernelSize(), l.getStride(), l.getPadding());
l.initializeConstraints();
return l;
}
@Override
public B kernelSize(int... kernelSize) {
super.kernelSize(ValidationUtils.validate2NonNegative(kernelSize, false, "kernelSize"));
return self();
}
@Override
public B stride(int... stride) {
super.stride(ValidationUtils.validate2NonNegative(stride, false, "stride"));
return self();
}
@Override
public B padding(int... padding) {
super.padding(ValidationUtils.validate2NonNegative(padding, false, "padding"));
return self();
}
@Override
public B dilation(int... dilation) {
super.dilation(ValidationUtils.validate2NonNegative(dilation, false, "dilation"));
return self();
}
}
}

View File

@ -45,15 +45,12 @@ import org.nd4j.linalg.learning.regularization.Regularization;
* the input activation. See {@link Dropout} for the full details
*/
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(builderMethodName = "innerBuilder")
public class DropoutLayer extends FeedForwardLayer {
{
setType(LayerType.DO);
}
public static DropoutLayerBuilder<?,?> builder() {
return innerBuilder();

View File

@ -36,6 +36,10 @@ import org.deeplearning4j.nn.conf.preprocessor.RnnToFeedForwardPreProcessor;
@EqualsAndHashCode(callSuper = true)
@SuperBuilder
public abstract class FeedForwardLayer extends BaseLayerConfiguration {
public static abstract class FeedForwardLayerBuilder<C extends FeedForwardLayer, B extends FeedForwardLayerBuilder<C, B>>
extends BaseLayerConfigurationBuilder<C, B> {
}
/**
* Number of inputs for the layer (usually the size of the last layer). <br> Note that for Convolutional layers,
* this is the input channels, otherwise is the previous layer size.
@ -55,7 +59,7 @@ public abstract class FeedForwardLayer extends BaseLayerConfiguration {
* this is the input channels, otherwise is the previous layer size.
*
*/
@Getter
@Getter @Setter
protected long nOut;
protected DataFormat timeDistributedFormat;

View File

@ -57,10 +57,10 @@ public abstract class LayerConfiguration
implements ILayerConfiguration, Serializable, Cloneable { // ITrainableLayerConfiguration
@Getter @Setter protected String name;
@Getter protected List<LayerConstraint> allParamConstraints;
@Getter protected List<LayerConstraint> weightConstraints;
@Getter protected List<LayerConstraint> biasConstraints;
@Getter protected List<LayerConstraint> constraints;
@Getter @Setter protected List<LayerConstraint> allParamConstraints;
@Getter @Setter protected List<LayerConstraint> weightConstraints;
@Getter @Setter protected List<LayerConstraint> biasConstraints;
@Getter @Setter protected List<LayerConstraint> constraints;
@Getter @Setter protected IWeightNoise weightNoise;
@Builder.Default private @Getter @Setter LinkedHashSet<String> variables = new LinkedHashSet<>();
@Getter @Setter private IDropout dropOut;
@ -325,4 +325,15 @@ public abstract class LayerConfiguration
runInheritance(getNetConfiguration());
}
public abstract static class LayerConfigurationBuilder<
C extends LayerConfiguration, B extends LayerConfigurationBuilder<C, B>> {
public B dropOut(double d) {
this.dropOut(new Dropout(d));
return self();
}
public B dropOut(IDropout d) {
this.dropOut = d;
return self();
}
}
}

View File

@ -61,9 +61,6 @@ public class LearnedSelfAttentionLayer extends SameDiffLayer {
/** Number of queries to learn */
private int nQueries;
private LearnedSelfAttentionLayer() {
/*No arg constructor for serialization*/
}
@Override
public InputPreProcessor getPreProcessorForInputType(InputType inputType) {

View File

@ -32,9 +32,6 @@ import org.nd4j.linalg.learning.regularization.Regularization;
@SuperBuilder
public abstract class NoParamLayer extends LayerConfiguration {
{
setType(LayerType.POOL);
}
@Override
public ParamInitializer initializer() {

View File

@ -42,7 +42,6 @@ import org.nd4j.linalg.api.ndarray.INDArray;
* filter/kernel size, the stride and padding The pooling layer takes the kernel size
*/
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@ -103,20 +102,20 @@ public class SeparableConvolution2D extends ConvolutionLayer {
}
public boolean hasBias() {
return hasBias;
return isHasBias();
}
@Override
public SeparableConvolution2D clone() {
SeparableConvolution2D clone = (SeparableConvolution2D) super.clone();
if (clone.kernelSize != null) {
clone.kernelSize = clone.kernelSize.clone();
if (clone.getKernelSize() != null) {
clone.setKernelSize( clone.getKernelSize().clone());
}
if (clone.stride != null) {
clone.stride = clone.stride.clone();
if (clone.getStride() != null) {
clone.setStride( clone.getStride().clone());
}
if (clone.padding != null) {
clone.padding = clone.padding.clone();
if (clone.getPadding() != null) {
clone.setPadding( clone.getPadding().clone());
}
return clone;
}
@ -165,11 +164,11 @@ public class SeparableConvolution2D extends ConvolutionLayer {
return InputTypeUtil.getOutputTypeCnnLayers(
inputType,
kernelSize,
stride,
padding,
dilation,
convolutionMode,
getKernelSize(),
getStride(),
getPadding(),
getDilation(),
getConvolutionMode(),
nOut,
layerIndex,
getName(),

View File

@ -20,6 +20,9 @@
package org.deeplearning4j.nn.conf.layers.misc;
import java.util.Collection;
import java.util.List;
import java.util.Set;
import lombok.EqualsAndHashCode;
import lombok.Getter;
import lombok.Setter;
@ -37,25 +40,20 @@ import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.learning.config.IUpdater;
import org.nd4j.linalg.learning.regularization.Regularization;
import com.fasterxml.jackson.annotation.JsonProperty;
import java.util.Collection;
import java.util.List;
import java.util.Set;
@EqualsAndHashCode(callSuper = false)
@SuperBuilder
@SuperBuilder(builderMethodName = "innerBuilder")
public class FrozenLayer extends LayerConfiguration {
/**
* A layer configuration, only if this layer config has been created from another one
*/
@Getter @Setter
private LayerConfiguration innerConfiguration;
/** A layer configuration, only if this layer config has been created from another one */
@Getter @Setter private LayerConfiguration innerConfiguration;
public static FrozenLayerBuilder<?, ?> builder() {
return innerBuilder();
}
public FrozenLayer(@JsonProperty("layer") LayerConfiguration layer) {
this.innerConfiguration = layer;
public static FrozenLayerBuilder<?, ?> builder(LayerConfiguration innerConfiguration) {
return innerBuilder().innerConfiguration(innerConfiguration);
}
@Override
@ -66,13 +64,23 @@ public class FrozenLayer extends LayerConfiguration {
}
@Override
public org.deeplearning4j.nn.api.Layer instantiate(NeuralNetConfiguration conf,
Collection<TrainingListener> trainingListeners, int layerIndex, INDArray layerParamsView,
boolean initializeParams, DataType networkDataType) {
public org.deeplearning4j.nn.api.Layer instantiate(
NeuralNetConfiguration conf,
Collection<TrainingListener> trainingListeners,
int layerIndex,
INDArray layerParamsView,
boolean initializeParams,
DataType networkDataType) {
//Need to be able to instantiate a layer, from a config - for JSON -> net type situations
org.deeplearning4j.nn.api.Layer underlying = innerConfiguration.instantiate(getNetConfiguration(), trainingListeners,
layerIndex, layerParamsView, initializeParams, networkDataType);
// Need to be able to instantiate a layer, from a config - for JSON -> net type situations
org.deeplearning4j.nn.api.Layer underlying =
innerConfiguration.instantiate(
getNetConfiguration(),
trainingListeners,
layerIndex,
layerParamsView,
initializeParams,
networkDataType);
NeuralNetConfiguration nncUnderlying = underlying.getNetConfiguration();
if (nncUnderlying.getNetWideVariables() != null) {
@ -109,7 +117,7 @@ public class FrozenLayer extends LayerConfiguration {
}
@Override
public List<Regularization> getRegularizationByParam(String param){
public List<Regularization> getRegularizationByParam(String param) {
return null;
}
@ -139,6 +147,4 @@ public class FrozenLayer extends LayerConfiguration {
this.constraints = constraints;
this.innerConfiguration.setConstraints(constraints);
}
}

View File

@ -22,6 +22,7 @@ package org.deeplearning4j.nn.conf.layers.misc;
import lombok.Data;
import lombok.EqualsAndHashCode;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.ParamInitializer;
import org.deeplearning4j.nn.api.layers.LayerConstraint;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
@ -39,19 +40,23 @@ import java.util.Collection;
import java.util.List;
import java.util.Set;
@Data
@EqualsAndHashCode(callSuper = false)
@SuperBuilder(builderMethodName = "innerBuilder")
public class FrozenLayerWithBackprop extends BaseWrapperLayerConfiguration {
public static FrozenLayerWithBackpropBuilder<?, ?> builder() {
return innerBuilder();
}
/**
* Create a new Frozen Layer, that wraps another layer with backpropagation enabled.
*
* @param layer configuration of the layer to wrap
* @param innerConfiguration configuration of the layer to wrap
*/
public FrozenLayerWithBackprop(@JsonProperty("layer") LayerConfiguration layer) {
super(layer);
public static FrozenLayerWithBackpropBuilder<?, ?> builder(LayerConfiguration innerConfiguration) {
return innerBuilder().underlying(innerConfiguration);
}
public NeuralNetConfiguration getInnerConf(NeuralNetConfiguration conf) {
NeuralNetConfiguration nnc = conf.clone();
nnc.getLayerConfigurations().add(0, underlying);

View File

@ -46,7 +46,7 @@ import org.nd4j.linalg.lossfunctions.ILossFunction;
import org.nd4j.linalg.lossfunctions.impl.LossL2;
import org.nd4j.serde.jackson.shaded.NDArrayTextSerializer;
@Data
@EqualsAndHashCode(callSuper = false)
@SuperBuilder(buildMethodName = "initBuild")
public class Yolo2OutputLayer extends LayerConfiguration {
@ -55,20 +55,20 @@ public class Yolo2OutputLayer extends LayerConfiguration {
* Loss function coefficient for position and size/scale components of the loss function. Default
* (as per paper): 5
*/
@Builder.Default private double lambdaCoord = 5;
@Builder.Default @Getter private double lambdaCoord = 5;
/**
* Loss function coefficient for the "no object confidence" components of the loss function.
* Default (as per paper): 0.5
*/
@Builder.Default private double lambdaNoObj = 0.5;
@Builder.Default @Getter private double lambdaNoObj = 0.5;
/** Loss function for position/scale component of the loss function */
@Builder.Default private ILossFunction lossPositionScale = new LossL2();
@Builder.Default @Getter private ILossFunction lossPositionScale = new LossL2();
/**
* Loss function for the class predictions - defaults to L2 loss (i.e., sum of squared errors, as
* per the paper), however Loss MCXENT could also be used (which is more common for
* classification).
*/
@Builder.Default private ILossFunction lossClassPredictions = new LossL2();
@Builder.Default @Getter private ILossFunction lossClassPredictions = new LossL2();
;
/**
* Bounding box priors dimensions [width, height]. For N bounding boxes, input has shape [rows,
@ -78,15 +78,12 @@ public class Yolo2OutputLayer extends LayerConfiguration {
*/
@JsonSerialize(using = NDArrayTextSerializer.class)
@JsonDeserialize(using = BoundingBoxesDeserializer.class)
@Builder.Default
@Builder.Default @Getter
private INDArray boundingBoxes;
@Builder.Default
@Builder.Default @Getter
private CNN2DFormat format = CNN2DFormat.NCHW; // Default for serialization of old formats
private Yolo2OutputLayer() {
// No-arg constructor for Jackson JSON
}
@Override
public Layer instantiate(

View File

@ -20,6 +20,7 @@
package org.deeplearning4j.nn.conf.layers.recurrent;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
import org.deeplearning4j.nn.conf.inputs.InputType;
import org.deeplearning4j.nn.conf.layers.LayerConfiguration;
@ -30,14 +31,18 @@ import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import java.util.Collection;
@SuperBuilder(builderMethodName = "innerBuilder")
public class LastTimeStep extends BaseWrapperLayerConfiguration {
private LastTimeStep() {}
public static LastTimeStepBuilder<?,?> builder() {
return innerBuilder();
}
public LastTimeStep(LayerConfiguration underlying) {
super(underlying);
this.name = underlying.getName(); // needed for keras import to match names
public static LastTimeStepBuilder<?,?> builder(LayerConfiguration underlying) {
return innerBuilder()
.underlying(underlying)
.name(underlying.getName());
}
public LayerConfiguration getUnderlying() {

View File

@ -41,7 +41,6 @@ import java.util.Map;
@EqualsAndHashCode(callSuper = false)
@NoArgsConstructor
@SuperBuilder
public class SimpleRnn extends BaseRecurrentLayer {
/**

View File

@ -20,9 +20,9 @@
package org.deeplearning4j.nn.conf.layers.recurrent;
import lombok.Data;
import lombok.EqualsAndHashCode;
import lombok.NonNull;
import java.util.Collection;
import lombok.*;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.conf.InputPreProcessor;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
import org.deeplearning4j.nn.conf.RNNFormat;
@ -33,55 +33,56 @@ import org.deeplearning4j.nn.layers.recurrent.TimeDistributedLayer;
import org.deeplearning4j.optimize.api.TrainingListener;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import com.fasterxml.jackson.annotation.JsonProperty;
import java.util.Collection;
@Data
@EqualsAndHashCode(callSuper = true)
@SuperBuilder
public class TimeDistributed extends BaseWrapperLayerConfiguration {
private RNNFormat rnnDataFormat = RNNFormat.NCW;
/**
* @param underlying Underlying (internal) layer - should be a feed forward type such as DenseLayerConfiguration
*/
public TimeDistributed(@JsonProperty("underlying") @NonNull LayerConfiguration underlying, @JsonProperty("rnnDataFormat") RNNFormat rnnDataFormat) {
super(underlying);
this.rnnDataFormat = rnnDataFormat;
}
public TimeDistributed(LayerConfiguration underlying){
super(underlying);
}
@Getter @Setter private RNNFormat rnnDataFormat = RNNFormat.NCW;
@Override
public org.deeplearning4j.nn.api.Layer instantiate(NeuralNetConfiguration conf, Collection<TrainingListener> trainingListeners,
int layerIndex, INDArray layerParamsView, boolean initializeParams, DataType networkDataType) {
public org.deeplearning4j.nn.api.Layer instantiate(
NeuralNetConfiguration conf,
Collection<TrainingListener> trainingListeners,
int layerIndex,
INDArray layerParamsView,
boolean initializeParams,
DataType networkDataType) {
LayerConfiguration lconf = conf.getFlattenedLayerConfigurations().get(layerIndex);
NeuralNetConfiguration conf2 = conf.clone();
conf2.setLayer(((TimeDistributed) lconf).getUnderlying());
return new TimeDistributedLayer(underlying.instantiate(conf2, trainingListeners, layerIndex, layerParamsView,
initializeParams, networkDataType), rnnDataFormat);
return new TimeDistributedLayer(
underlying.instantiate(
conf2,
trainingListeners,
layerIndex,
layerParamsView,
initializeParams,
networkDataType),
rnnDataFormat);
}
@Override
public InputType getOutputType(int layerIndex, InputType inputType) {
if (inputType.getType() != InputType.Type.RNN) {
throw new IllegalStateException("Only RNN input type is supported as input to TimeDistributed layer (layer #" + layerIndex + ")");
throw new IllegalStateException(
"Only RNN input type is supported as input to TimeDistributed layer (layer #"
+ layerIndex
+ ")");
}
InputType.InputTypeRecurrent rnn = (InputType.InputTypeRecurrent) inputType;
InputType ff = InputType.feedForward(rnn.getSize());
InputType ffOut = underlying.getOutputType(layerIndex, ff);
return InputType.recurrent(ffOut.arrayElementsPerExample(), rnn.getTimeSeriesLength(), rnnDataFormat);
return InputType.recurrent(
ffOut.arrayElementsPerExample(), rnn.getTimeSeriesLength(), rnnDataFormat);
}
@Override
public void setNIn(InputType inputType, boolean override) {
if (inputType.getType() != InputType.Type.RNN) {
throw new IllegalStateException("Only RNN input type is supported as input to TimeDistributed layer");
throw new IllegalStateException(
"Only RNN input type is supported as input to TimeDistributed layer");
}
InputType.InputTypeRecurrent rnn = (InputType.InputTypeRecurrent) inputType;
@ -92,7 +93,7 @@ public class TimeDistributed extends BaseWrapperLayerConfiguration {
@Override
public InputPreProcessor getPreProcessorForInputType(InputType inputType) {
//No preprocessor - the wrapper layer operates as the preprocessor
// No preprocessor - the wrapper layer operates as the preprocessor
return null;
}
}

View File

@ -20,6 +20,7 @@
package org.deeplearning4j.nn.conf.layers.samediff;
import lombok.Builder;
import lombok.EqualsAndHashCode;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.Layer;
@ -47,7 +48,9 @@ public abstract class SameDiffLayer extends AbstractSameDiffLayer {
/**
* WeightInit, default is XAVIER.
*/
@Builder.Default
protected WeightInit weightInit = WeightInit.XAVIER;
@Builder.Default
protected Map<String,IWeightInit> paramWeightInit = new HashMap<>();

View File

@ -20,6 +20,7 @@
package org.deeplearning4j.nn.conf.layers.samediff;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
import org.deeplearning4j.nn.conf.layers.LayerConfiguration;
import org.deeplearning4j.optimize.api.TrainingListener;
@ -30,13 +31,10 @@ import org.nd4j.linalg.api.ndarray.INDArray;
import java.util.Collection;
import java.util.Map;
@SuperBuilder
public abstract class SameDiffOutputLayer extends AbstractSameDiffLayer {
protected SameDiffOutputLayer() {
//No op constructor for Jackson
}
/**
* Define the output layer

View File

@ -21,6 +21,7 @@
package org.deeplearning4j.nn.conf.layers.util;
import lombok.NoArgsConstructor;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.ParamInitializer;
import org.deeplearning4j.nn.conf.InputPreProcessor;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
@ -38,7 +39,7 @@ import java.util.Collection;
import java.util.List;
import java.util.Map;
@NoArgsConstructor
@SuperBuilder
public class MaskLayer extends NoParamLayer {
@Override
public org.deeplearning4j.nn.api.Layer instantiate(NeuralNetConfiguration conf,

View File

@ -35,22 +35,17 @@ import com.fasterxml.jackson.annotation.JsonProperty;
import java.util.Collection;
@Data
@EqualsAndHashCode(callSuper = false)
@SuperBuilder
public class MaskZeroLayer extends BaseWrapperLayerConfiguration {
@Builder.Default
@Builder.Default @Getter @Setter
private double maskingValue = 0.0;
private static final long serialVersionUID = 9074525846200921839L;
public MaskZeroLayer(@JsonProperty("underlying") LayerConfiguration underlying, @JsonProperty("maskingValue") double maskingValue) {
this.underlying = underlying;
this.maskingValue = maskingValue;
}
@Override
public org.deeplearning4j.nn.api.Layer instantiate(NeuralNetConfiguration conf,

View File

@ -23,6 +23,7 @@ package org.deeplearning4j.nn.conf.layers.wrapper;
import java.util.List;
import lombok.EqualsAndHashCode;
import lombok.Getter;
import lombok.Setter;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.ParamInitializer;
import org.deeplearning4j.nn.conf.InputPreProcessor;
@ -42,7 +43,8 @@ import org.nd4j.linalg.learning.regularization.Regularization;
public abstract class BaseWrapperLayerConfiguration extends LayerConfiguration {
/** The configuration to of another layer to wrap */
@Getter protected LayerConfiguration underlying;
@Getter @Setter
protected LayerConfiguration underlying;
/**
* Set the net configuration for this configuration as well as for the underlying layer (if not

View File

@ -38,8 +38,6 @@ import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.learning.regularization.Regularization;
@Data
@NoArgsConstructor
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@JsonIgnoreProperties("lossFn")

View File

@ -349,6 +349,6 @@ public abstract class BaseOutputLayer<LayerConfT extends org.deeplearning4j.nn.c
@Override
public boolean hasBias() {
return getTypedLayerConfiguration().hasBias();
return getTypedLayerConfiguration().isHasBias();
}
}

View File

@ -226,7 +226,7 @@ public class DefaultParamInitializer extends AbstractParamInitializer {
protected boolean hasBias(LayerConfiguration layer){
if(layer instanceof BaseOutputLayer ) {
return ((BaseOutputLayer) layer).hasBias();
return ((BaseOutputLayer) layer).isHasBias();
} else if(layer instanceof DenseLayer){
return ((DenseLayer)layer).isHasBias();
} else if(layer instanceof EmbeddingLayer){

View File

@ -382,7 +382,7 @@ public class TransferLearning {
}
LayerConfiguration origLayerConf = editedModel.getNetConfiguration().getFlattenedLayerConfigurations().get(i);
LayerConfiguration newLayerConf = new org.deeplearning4j.nn.conf.layers.misc.FrozenLayer(origLayerConf);
LayerConfiguration newLayerConf = org.deeplearning4j.nn.conf.layers.misc.FrozenLayer.builder().innerConfiguration(origLayerConf).build();
newLayerConf.setName(origLayerConf.getName());
editedModel.getNetConfiguration().getNetConfigurations().get(i).setLayer(newLayerConf);
}
@ -1009,7 +1009,7 @@ public class TransferLearning {
String layerName = gv.getVertexName();
LayerVertex currLayerVertex = (LayerVertex) newConfig.getVertices().get(layerName);
LayerConfiguration origLayerConf = currLayerVertex.getLayerConfiguration();
LayerConfiguration newLayerConf = new org.deeplearning4j.nn.conf.layers.misc.FrozenLayer(origLayerConf);
LayerConfiguration newLayerConf = org.deeplearning4j.nn.conf.layers.misc.FrozenLayer.builder().innerConfiguration(origLayerConf).build();
newLayerConf.setName(origLayerConf.getName());
//Complication here(and reason for clone on next line): inner LayerConfiguration (implementation)
// NeuralNetConfiguration.layer (config) should keep the original layer config. While network