Brutex
/
cavis
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Using @SuperBuilder for LayerConfigurations 

Signed-off-by: brian <brian@brutex.de>
master
Brian Rosenberger 2023-04-27 15:48:34 +02:00
parent 396dbec24e
commit 7628bbdd53
37 changed files with 938 additions and 808 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/ActivationLayer.java
View File

@ -43,7 +43,7 @@ import org.nd4j.linalg.learning.config.IUpdater;
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class ActivationLayer extends NoParamLayer {
@ -133,8 +133,12 @@ public class ActivationLayer extends NoParamLayer {
public static abstract class ActivationLayerBuilder<
C extends ActivationLayer, B extends ActivationLayerBuilder<C, B>>
extends NoParamLayer.NoParamLayerBuilder<C, B> {
public C build() {
C l = this.initBuild();
}
private static final class ActivationLayerBuilderImpl extends ActivationLayerBuilder<ActivationLayer, ActivationLayerBuilderImpl> {
public ActivationLayer build() {
ActivationLayer l = this.initBuild();
l.initializeConstraints();
return l;
}

2
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/BaseUpsamplingLayer.java
View File

@ -33,7 +33,7 @@ import org.deeplearning4j.nn.conf.inputs.InputType;
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder()
@SuperBuilder
public abstract class BaseUpsamplingLayer extends NoParamLayer {
/**

183
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/BatchNormalization.java
View File

@ -25,10 +25,10 @@ import java.util.List;
import java.util.Map;
import lombok.*;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.layers.LayerConstraint;
import net.brutex.ai.dnn.api.LayerType;
import org.deeplearning4j.nn.api.Layer;
import org.deeplearning4j.nn.api.ParamInitializer;
import org.deeplearning4j.nn.api.layers.LayerConstraint;
import org.deeplearning4j.nn.conf.CNN2DFormat;
import org.deeplearning4j.nn.conf.InputPreProcessor;
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
@ -48,101 +48,117 @@ import org.nd4j.linalg.learning.regularization.Regularization;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class BatchNormalization extends FeedForwardLayer {
/**
* At test time: we can use a global estimate of the mean and variance, calculated using a moving average of the
* batch means/variances. This moving average is implemented as:<br> globalMeanEstimate = decay *
* globalMeanEstimate + (1-decay) * batchMean<br> globalVarianceEstimate = decay * globalVarianceEstimate +
* (1-decay) * batchVariance<br>
* At test time: we can use a global estimate of the mean and variance, calculated using a moving
* average of the batch means/variances. This moving average is implemented as:<br>
* globalMeanEstimate = decay * globalMeanEstimate + (1-decay) * batchMean<br>
* globalVarianceEstimate = decay * globalVarianceEstimate + (1-decay) * batchVariance<br>
*
* @param decay Decay value to use for global stats calculation
*/
@lombok.Builder.Default
protected double decay = 0.9;
@lombok.Builder.Default protected double decay = 0.9;
// Note: need to set defaults here in addition to builder, in case user uses no-op constructor...
/**
* Epsilon value for batch normalization; small floating point value added to variance (algorithm 1 in <a
* href="https://arxiv.org/pdf/1502.03167v3.pdf">https://arxiv.org/pdf/1502.03167v3.pdf</a>) to reduce/avoid
* underflow issues.<br> Default: 1e-5
* Epsilon value for batch normalization; small floating point value added to variance (algorithm
* 1 in <a
* href="https://arxiv.org/pdf/1502.03167v3.pdf">https://arxiv.org/pdf/1502.03167v3.pdf</a>) to
* reduce/avoid underflow issues.<br>
* Default: 1e-5
*
* @param eps Epsilon values to use
*/
@lombok.Builder.Default protected double eps = 1e-5;
/**
* If doing minibatch training or not. Default: true. Under most circumstances, this should be set to true. If
* doing full batch training (i.e., all examples in a single DataSet object - very small data sets) then this
* should be set to false. Affects how global mean/variance estimates are calculated.
* If doing minibatch training or not. Default: true. Under most circumstances, this should be set
* to true. If doing full batch training (i.e., all examples in a single DataSet object - very
* small data sets) then this should be set to false. Affects how global mean/variance estimates
* are calculated.
*
* @param minibatch Minibatch parameter
*/
@lombok.Builder.Default protected boolean isMinibatch = true;
/**
* Used only when 'true' is passed to {@link BatchNormalizationBuilder#lockGammaBeta(boolean)}. Value is not used otherwise.<br> Default:
* 1.0
* Used only when 'true' is passed to {@link BatchNormalizationBuilder#lockGammaBeta(boolean)}.
* Value is not used otherwise.<br>
* Default: 1.0
*
* @param gamma Gamma parameter for all activations, used only with locked gamma/beta configuration mode
* @param gamma Gamma parameter for all activations, used only with locked gamma/beta
* configuration mode
*/
@lombok.Builder.Default protected double gamma = 1.0;
/**
* Used only when 'true' is passed to {@link BatchNormalizationBuilder#lockGammaBeta(boolean)}. Value is not used otherwise.<br> Default:
* 0.0
* Used only when 'true' is passed to {@link BatchNormalizationBuilder#lockGammaBeta(boolean)}.
* Value is not used otherwise.<br>
* Default: 0.0
*
* @param beta Beta parameter for all activations, used only with locked gamma/beta configuration mode
* @param beta Beta parameter for all activations, used only with locked gamma/beta configuration
* mode
*/
@lombok.Builder.Default protected double beta = 0.0;
/**
* Set constraints to be applied to the beta parameter of this batch normalisation layer. Default: no
* constraints.<br> Constraints can be used to enforce certain conditions (non-negativity of parameters,
* max-norm regularization, etc). These constraints are applied at each iteration, after the parameters have
* been updated.
*
* Set constraints to be applied to the beta parameter of this batch normalisation layer. Default:
* no constraints.<br>
* Constraints can be used to enforce certain conditions (non-negativity of parameters, max-norm
* regularization, etc). These constraints are applied at each iteration, after the parameters
* have been updated.
*/
protected List<LayerConstraint> betaConstraints;
/**
* Set constraints to be applied to the gamma parameter of this batch normalisation layer. Default: no
* constraints.<br> Constraints can be used to enforce certain conditions (non-negativity of parameters,
* max-norm regularization, etc). These constraints are applied at each iteration, after the parameters have
* been updated.
*
* Set constraints to be applied to the gamma parameter of this batch normalisation layer.
* Default: no constraints.<br>
* Constraints can be used to enforce certain conditions (non-negativity of parameters, max-norm
* regularization, etc). These constraints are applied at each iteration, after the parameters
* have been updated.
*/
protected List<LayerConstraint> gammaConstraints;
/**
* When using CuDNN or MKLDNN and an error is encountered, should fallback to the non-helper implementation be allowed?
* If set to false, an exception in the helper will be propagated back to the user. If true, the built-in
* (non-MKL/CuDNN) implementation for BatchNormalizationLayer will be used
* When using CuDNN or MKLDNN and an error is encountered, should fallback to the non-helper
* implementation be allowed? If set to false, an exception in the helper will be propagated back
* to the user. If true, the built-in (non-MKL/CuDNN) implementation for BatchNormalizationLayer
* will be used
*
* @param allowFallback Whether fallback to non-CuDNN implementation should be used
*/
@lombok.Builder.Default protected boolean cudnnAllowFallback = true;
/**
* How should the moving average of variance be stored? Two different parameterizations are supported.
* useLogStd(false): equivalent to 1.0.0-beta3 and earlier. The variance "parameter" is stored directly as
* variable<br> useLogStd(true): (Default) variance is stored as log10(stdev)<br> The motivation here is for
* numerical stability (FP16 etc) and also distributed training: storing the variance directly can cause
* numerical issues. For example, a standard deviation of 1e-3 (something that could be encountered in practice)
* gives a variance of 1e-6, which can be problematic for 16-bit floating point
* How should the moving average of variance be stored? Two different parameterizations are
* supported. useLogStd(false): equivalent to 1.0.0-beta3 and earlier. The variance "parameter" is
* stored directly as variable<br>
* useLogStd(true): (Default) variance is stored as log10(stdev)<br>
* The motivation here is for numerical stability (FP16 etc) and also distributed training:
* storing the variance directly can cause numerical issues. For example, a standard deviation of
* 1e-3 (something that could be encountered in practice) gives a variance of 1e-6, which can be
* problematic for 16-bit floating point
*
* How should the moving average of variance be stored? Two different parameterizations are supported.
* useLogStd(false): equivalent to 1.0.0-beta3 and earlier. The variance "parameter" is stored directly as
* variable<br> useLogStd(true): (Default) variance is stored as log10(stdev)<br> The motivation here is for
* numerical stability (FP16 etc) and also distributed training: storing the variance directly can cause
* numerical issues. For example, a standard deviation of 1e-3 (something that could be encountered in practice)
* gives a variance of 1e-6, which can be problematic for 16-bit floating point
* <p>How should the moving average of variance be stored? Two different parameterizations are
* supported. useLogStd(false): equivalent to 1.0.0-beta3 and earlier. The variance "parameter" is
* stored directly as variable<br>
* useLogStd(true): (Default) variance is stored as log10(stdev)<br>
* The motivation here is for numerical stability (FP16 etc) and also distributed training:
* storing the variance directly can cause numerical issues. For example, a standard deviation of
* 1e-3 (something that could be encountered in practice) gives a variance of 1e-6, which can be
* problematic for 16-bit floating point
*/
@lombok.Builder.Default protected boolean useLogStd = false; //Default for deserialized models (1.0.0-beta3) and earlier: store variance as variance. Post 1.0.0-beta3: use log stdev instead
@lombok.Builder.Default
protected boolean useLogStd =
false; // Default for deserialized models (1.0.0-beta3) and earlier: store variance as
// variance. Post 1.0.0-beta3: use log stdev instead
/**
* Set the input and output array data format. Defaults to NCHW format - i.e., channels first.
* See {@link CNN2DFormat} for more details
* Set the input and output array data format. Defaults to NCHW format - i.e., channels first. See
* {@link CNN2DFormat} for more details
*
* @param format Format to use
*/
@lombok.Builder.Default protected CNN2DFormat dataFormat = CNN2DFormat.NCHW; //Default for deserialized models, 1.0.0-beta6 and earlier
@lombok.Builder.Default
protected CNN2DFormat dataFormat =
CNN2DFormat.NCHW; // Default for deserialized models, 1.0.0-beta6 and earlier
private boolean lockGammaBeta;
@ -151,14 +167,11 @@ public class BatchNormalization extends FeedForwardLayer {
}
public static BatchNormalizationBuilder<?, ?> builder(double gamma, double beta) {
return innerBuilder()
.gamma(gamma)
.beta(beta);
return innerBuilder().gamma(gamma).beta(beta);
}
public static BatchNormalizationBuilder<?, ?> builder(boolean lockGammaBeta) {
return innerBuilder()
.lockGammaBeta(lockGammaBeta);
return innerBuilder().lockGammaBeta(lockGammaBeta);
}
@Override
@ -168,8 +181,13 @@ public class BatchNormalization extends FeedForwardLayer {
}
@Override
public Layer instantiate(NeuralNetConfiguration conf, Collection<TrainingListener> trainingListeners,
int layerIndex, INDArray layerParamsView, boolean initializeParams, DataType networkDataType) {
public Layer instantiate(
NeuralNetConfiguration conf,
Collection<TrainingListener> trainingListeners,
int layerIndex,
INDArray layerParamsView,
boolean initializeParams,
DataType networkDataType) {
this.setNetConfiguration(conf);
LayerValidation.assertNOutSet("BatchNormalization", getName(), layerIndex, getNOut());
runInheritance();
@ -196,7 +214,8 @@ public class BatchNormalization extends FeedForwardLayer {
if (inputType == null) {
throw new IllegalStateException(
"Invalid input type: Batch norm layer expected input of type CNN, got null for layer \""
+ getName() + "\"");
+ getName()
+ "\"");
}
// Can handle CNN, flat CNN, CNN3D or FF input formats only
@ -209,7 +228,10 @@ public class BatchNormalization extends FeedForwardLayer {
default:
throw new IllegalStateException(
"Invalid input type: Batch norm layer expected input of type CNN, CNN Flat or FF, got "
+ inputType + " for layer index " + layerIndex + ", layer name = "
+ inputType
+ " for layer index "
+ layerIndex
+ ", layer name = "
+ getName());
}
}
@ -233,7 +255,10 @@ public class BatchNormalization extends FeedForwardLayer {
default:
throw new IllegalStateException(
"Invalid input type: Batch norm layer expected input of type CNN, CNN Flat or FF, got "
+ inputType + " for layer " + getName() + "\"");
+ inputType
+ " for layer "
+ getName()
+ "\"");
}
nOut = nIn;
}
@ -253,7 +278,8 @@ public class BatchNormalization extends FeedForwardLayer {
@Override
public List<Regularization> getRegularizationByParam(String paramName) {
//Don't regularize batch norm params: similar to biases in the sense that there are not many of them...
// Don't regularize batch norm params: similar to biases in the sense that there are not many of
// them...
return null;
}
@ -288,16 +314,24 @@ public class BatchNormalization extends FeedForwardLayer {
// During forward pass: working memory size approx. equal to 2x input size (copy ops, etc)
val inferenceWorkingSize = 2 * inputType.arrayElementsPerExample();
//During training: we calculate mean and variance... result is equal to nOut, and INDEPENDENT of minibatch size
// During training: we calculate mean and variance... result is equal to nOut, and INDEPENDENT
// of minibatch size
val trainWorkFixed = 2 * nOut;
//During backprop: multiple working arrays... output size, 2 * output size (indep. of example size),
val trainWorkingSizePerExample = inferenceWorkingSize //Inference during backprop
// During backprop: multiple working arrays... output size, 2 * output size (indep. of example
// size),
val trainWorkingSizePerExample =
inferenceWorkingSize // Inference during backprop
+ (outputType.arrayElementsPerExample() + 2 * nOut); // Backprop gradient calculation
return new LayerMemoryReport.Builder(name, BatchNormalization.class, inputType, outputType)
.standardMemory(numParams, updaterStateSize)
.workingMemory(0, 0, trainWorkFixed, trainWorkingSizePerExample) //No additional memory (beyond activations) for inference
.cacheMemory(MemoryReport.CACHE_MODE_ALL_ZEROS, MemoryReport.CACHE_MODE_ALL_ZEROS) //No caching
.workingMemory(
0,
0,
trainWorkFixed,
trainWorkingSizePerExample) // No additional memory (beyond activations) for inference
.cacheMemory(
MemoryReport.CACHE_MODE_ALL_ZEROS, MemoryReport.CACHE_MODE_ALL_ZEROS) // No caching
.build();
}
@ -306,13 +340,20 @@ public class BatchNormalization extends FeedForwardLayer {
return false; // No pretrain params in BN
}
public static abstract class BatchNormalizationBuilder<C extends BatchNormalization, B extends BatchNormalizationBuilder<C, B>> extends FeedForwardLayerBuilder<C, B> {
public C build() {
C l = this.initBuild();
private static final class BatchNormalizationBuilderImpl
extends BatchNormalizationBuilder<BatchNormalization, BatchNormalizationBuilderImpl> {
public BatchNormalization build() {
BatchNormalization l = new BatchNormalization(this);
l.setType(LayerType.BN);
l.initializeConstraints();
return l;
}
}
public abstract static class BatchNormalizationBuilder<
C extends BatchNormalization, B extends BatchNormalizationBuilder<C, B>>
extends FeedForwardLayerBuilder<C, B> {
public B helperAllowFallback(boolean b) {
this.cudnnAllowFallback$value = b;
this.cudnnAllowFallback$set = true;
@ -323,12 +364,10 @@ public class BatchNormalization extends FeedForwardLayer {
this.betaConstraints = List.of(constraints);
return self();
}
public B constrainGamma(LayerConstraint... constraints) {
this.gammaConstraints = List.of(constraints);
return self();
}
}
}

49
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/CapsuleLayer.java
View File

@ -38,9 +38,8 @@ import org.nd4j.linalg.api.memory.MemoryWorkspace;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.factory.Nd4j;
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class CapsuleLayer extends SameDiffLayer {
private static final String WEIGHT_PARAM = "weight";
@ -78,6 +77,18 @@ public class CapsuleLayer extends SameDiffLayer {
*/
@Builder.Default @Getter @Setter private int routings = 3;
public static CapsuleLayerBuilder<?,?> builder() {
return innerBuilder()
;
}
public static CapsuleLayerBuilder<?,?> builder(int capsules, int capsulesDim, int routings) {
return innerBuilder()
.capsules(capsules)
.capsuleDimensions(capsulesDim)
.routings(routings);
}
@Override
public void setNIn(InputType inputType, boolean override) {
if(inputType == null || inputType.getType() != Type.RNN) {
@ -185,16 +196,6 @@ public class CapsuleLayer extends SameDiffLayer {
return InputType.recurrent(capsules, capsuleDimensions);
}
public static CapsuleLayerBuilder<?,?> builder() {
return innerBuilder()
;
}
public static CapsuleLayerBuilder<?,?> builder(int capsules, int capsulesDim, int routings) {
return innerBuilder()
.capsules(capsules)
.capsuleDimensions(capsulesDim)
.routings(routings);
}
public static abstract class CapsuleLayerBuilder<
C extends CapsuleLayer, B extends CapsuleLayerBuilder<C, B>>
extends SameDiffLayerBuilder<C, B> {
@ -215,35 +216,37 @@ public class CapsuleLayer extends SameDiffLayer {
}
public C build() {
C l = this.initBuild();
if (capsules <= 0 || capsuleDimensions <= 0 || routings$value <= 0) {
}
private static final class CapsuleLayerBuilderImpl extends CapsuleLayerBuilder<CapsuleLayer, CapsuleLayerBuilderImpl> {
public CapsuleLayer build() {
CapsuleLayer l = new CapsuleLayer(this);
if (l.getCapsules() <= 0 || l.getCapsuleDimensions() <= 0 || l.getRoutings() <= 0) {
throw new IllegalArgumentException(
"Invalid configuration for Capsule ILayer (layer name = \""
+ l.getName()
+ "\"):"
+ " capsules, capsuleDimensions, and routings must be > 0. Got: "
+ capsules
+ l.getCapsules()
+ ", "
+ capsuleDimensions
+ l.getCapsuleDimensions()
+ ", "
+ routings$value);
+ l.getRoutings());
}
if (inputCapsules$value < 0 || inputCapsuleDimensions$value < 0) {
if (l.getInputCapsules() < 0 || l.getInputCapsuleDimensions() < 0) {
throw new IllegalArgumentException(
"Invalid configuration for Capsule ILayer (layer name = \""
+ l.getName()
+ "\"):"
+ " inputCapsules and inputCapsuleDimensions must be >= 0 if set. Got: "
+ inputCapsules$value
+ l.getInputCapsules()
+ ", "
+ inputCapsuleDimensions$value);
+ l.getInputCapsuleDimensions() );
}
return l;
}
}
}

38
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/CenterLossOutputLayer.java
View File

@ -20,6 +20,8 @@
package org.deeplearning4j.nn.conf.layers;
import java.util.Collection;
import java.util.Map;
import lombok.*;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.Layer;
@ -30,36 +32,21 @@ import org.deeplearning4j.nn.conf.memory.LayerMemoryReport;
import org.deeplearning4j.nn.conf.memory.MemoryReport;
import org.deeplearning4j.nn.params.CenterLossParamInitializer;
import org.deeplearning4j.optimize.api.TrainingListener;
import org.nd4j.linalg.activations.impl.ActivationSoftmax;
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.config.NoOp;
import org.nd4j.linalg.lossfunctions.ILossFunction;
import org.nd4j.linalg.lossfunctions.LossFunctions.LossFunction;
import java.util.Collection;
import java.util.Map;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class CenterLossOutputLayer extends BaseOutputLayer {
@Builder.Default protected double alpha= 0.805;
@Builder.Default protected double lambda = 2e-4;
@Builder.Default protected boolean gradientCheck = false;
public static abstract class CenterLossOutputLayerBuilder<C extends CenterLossOutputLayer, B extends CenterLossOutputLayerBuilder<C,B>> extends
BaseOutputLayerBuilder<C, B> {
public C build() {
C l = initBuild();
l.initializeConstraints();
return l;
}
}
@Override
public Layer instantiate(NeuralNetConfiguration conf, Collection<TrainingListener> trainingListeners,
int layerIndex, INDArray layerParamsView, boolean initializeParams, DataType networkDataType) {
@ -91,7 +78,6 @@ public static abstract class CenterLossOutputLayerBuilder<C extends CenterLossOu
return getUpdater();
}
public boolean getGradientCheck() {
return gradientCheck;
}
@ -135,6 +121,24 @@ public static abstract class CenterLossOutputLayerBuilder<C extends CenterLossOu
.build();
}
public static abstract class CenterLossOutputLayerBuilder<C extends CenterLossOutputLayer, B extends CenterLossOutputLayerBuilder<C,B>> extends
BaseOutputLayerBuilder<C, B> {
public C build() {
C l = initBuild();
l.initializeConstraints();
return l;
}
}
private static final class CenterLossOutputLayerBuilderImpl extends CenterLossOutputLayerBuilder<CenterLossOutputLayer,
CenterLossOutputLayerBuilderImpl> {
public CenterLossOutputLayer build() {
CenterLossOutputLayer l = new CenterLossOutputLayer(this);
l.initializeConstraints();
return l;
}
}
}

2
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Convolution1D.java
View File

@ -29,6 +29,6 @@ import lombok.experimental.SuperBuilder;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class Convolution1D extends Convolution1DLayer {
}

21
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Convolution1DLayer.java
View File

@ -47,9 +47,8 @@ import org.nd4j.linalg.api.ndarray.INDArray;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class Convolution1DLayer extends ConvolutionLayer {
@Builder.Default private RNNFormat rnnDataFormat = RNNFormat.NCW;
/**
* Set the data format for the CNN activations - NCHW (channels first) or NHWC (channels last).
* See {@link CNN2DFormat} for more details.<br>
@ -60,6 +59,7 @@ public class Convolution1DLayer extends ConvolutionLayer {
@Builder.Default
protected CNN2DFormat dataFormat =
CNN2DFormat.NCHW; // default value for legacy serialization reasons
@Builder.Default private RNNFormat rnnDataFormat = RNNFormat.NCW;
/**
* Size of the convolution
*
@ -183,17 +183,20 @@ public class Convolution1DLayer extends ConvolutionLayer {
return true;
}
public static abstract class Convolution1DLayerBuilder<
C extends ConvolutionLayer, B extends Convolution1DLayerBuilder<C, B>>
extends ConvolutionLayerBuilder<C, B> {
public C build() {
C l = initBuild();
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), padding$value);
private static final class Convolution1DLayerBuilderImpl extends ConvolutionLayerBuilder<ConvolutionLayer, Convolution1DLayerBuilderImpl> {
public ConvolutionLayer build() {
ConvolutionLayer l = initBuild();
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), l.getPadding());
ConvolutionUtils.validateCnnKernelStridePadding(
kernelSize$value, stride$value, padding$value);
l.getKernelSize(), l.getStride(), l.getPadding());
l.initializeConstraints();
return l;
}
}
public static abstract class Convolution1DLayerBuilder<
C extends ConvolutionLayer, B extends Convolution1DLayerBuilder<C, B>>
extends ConvolutionLayerBuilder<C, B> {
public B kernelSize(int @NonNull ... kernelSize) {
this.kernelSize$value[0] = ValidationUtils.validate1NonNegative(kernelSize, "kernelSize")[0];

2
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Convolution2D.java
View File

@ -30,6 +30,6 @@ import lombok.experimental.SuperBuilder;
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class Convolution2D extends ConvolutionLayer {
}

17
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Convolution3D.java
View File

@ -40,7 +40,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(builderMethodName = "innerBuilder", buildMethodName = "initBuild")
@SuperBuilder(builderMethodName = "innerBuilder")
public class Convolution3D extends ConvolutionLayer {
/**
@ -235,17 +235,20 @@ public class Convolution3D extends ConvolutionLayer {
NDHWC
}
private static final class Convolution3DBuilderImpl extends Convolution3DBuilder<Convolution3D, Convolution3DBuilderImpl> {
public Convolution3D build() {
Convolution3D l = new Convolution3D(this);
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), l.getPadding());
Convolution3DUtils.validateCnn3DKernelStridePadding(l.getKernelSize(), l.getStride(), l.getPadding());
return l;
}
}
// public Builder(int[] kernelSize, int[] stride, int[] padding, int[] dilation) {
// sup/er(kernelSize, stride, padding, dilation, 3);
public static abstract class Convolution3DBuilder<
C extends Convolution3D, B extends Convolution3DBuilder<C, B>>
extends ConvolutionLayer.ConvolutionLayerBuilder<C, B> {
public C build() {
ConvolutionUtils.validateConvolutionModePadding(convolutionMode$value, padding);
Convolution3DUtils.validateCnn3DKernelStridePadding(kernelSize, stride, padding);
C l = initBuild();
return l;
}
@Override // TODO we can use the parent builder and do not need to redefine the variables.
// Validation can be done in override function!

86
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/ConvolutionLayer.java
View File

@ -48,7 +48,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
*/
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(builderMethodName = "innerBuilder", buildMethodName = "initBuild")
@SuperBuilder(builderMethodName = "innerBuilder")
public class ConvolutionLayer extends FeedForwardLayer {
/**
* Size of the convolution rows/columns
@ -397,48 +397,7 @@ public class ConvolutionLayer extends FeedForwardLayer {
return self();
}
public C build() {
ConvolutionUtils.validateConvolutionModePadding(convolutionMode$value, padding$value);
ConvolutionUtils.validateCnnKernelStridePadding(
kernelSize$value, stride$value, padding$value);
if (kernelSize$value.length != convolutionDim$value) {
throw new IllegalArgumentException(
"Kernel argument should be a "
+ convolutionDim$value
+ "d array, got "
+ Arrays.toString(kernelSize$value));
}
if (stride$value.length != convolutionDim$value) {
throw new IllegalArgumentException(
"Strides argument should be a "
+ convolutionDim$value
+ "d array, got "
+ Arrays.toString(stride$value));
}
if (padding$value.length != convolutionDim$value) {
throw new IllegalArgumentException(
"Padding argument should be a "
+ convolutionDim$value
+ "d array, got "
+ Arrays.toString(padding$value));
}
if (dilation$value.length != convolutionDim$value) {
throw new IllegalArgumentException(
"Dilation argument should be a "
+ convolutionDim$value
+ "d array, got "
+ Arrays.toString(dilation$value));
}
C l = initBuild();
l.setType(LayerType.CONV);
l.initializeConstraints();
return l;
}
/**
* When using CuDNN or MKLDNN and an error is encountered, should fallback to the non-helper
@ -454,4 +413,47 @@ public class ConvolutionLayer extends FeedForwardLayer {
return self();
}
}
private static final class ConvolutionLayerBuilderImpl extends ConvolutionLayerBuilder<ConvolutionLayer, ConvolutionLayerBuilderImpl> {
public ConvolutionLayer build() {
ConvolutionLayer l = new ConvolutionLayer(this);
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), l.getPadding());
ConvolutionUtils.validateCnnKernelStridePadding(
l.getKernelSize(), l.getStride(), l.getPadding());
if (l.getKernelSize().length != l.getConvolutionDim()) {
throw new IllegalArgumentException(
"Kernel argument should be a "
+ l.getConvolutionDim()
+ "d array, got "
+ Arrays.toString(l.getKernelSize()));
}
if (l.getStride().length != l.getConvolutionDim()) {
throw new IllegalArgumentException(
"Strides argument should be a "
+ l.getConvolutionDim()
+ "d array, got "
+ Arrays.toString(l.getStride()));
}
if (l.getPadding().length != l.getConvolutionDim()) {
throw new IllegalArgumentException(
"Padding argument should be a "
+ l.getConvolutionDim()
+ "d array, got "
+ Arrays.toString(l.getPadding()));
}
if (l.getDilation().length != l.getConvolutionDim()) {
throw new IllegalArgumentException(
"Dilation argument should be a "
+ l.getConvolutionDim()
+ "d array, got "
+ Arrays.toString(l.getDilation()));
}
l.setType(LayerType.CONV);
l.initializeConstraints();
return l;
}
}
}

11
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Deconvolution2D.java
View File

@ -46,7 +46,7 @@ import java.util.Map;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuild")
@SuperBuilder(builderMethodName = "innerBuilder")
public class Deconvolution2D extends ConvolutionLayer {
@ -57,12 +57,15 @@ private CNN2DFormat format = CNN2DFormat.NCHW;
return false;
}
public static abstract class Deconvolution2DBuilder<C extends Deconvolution2D, B extends Deconvolution2DBuilder<C, B>> extends ConvolutionLayerBuilder<C, B> {
public C build() {
C l = initBuild();
private static final class Deconvolution2DBuilderImpl extends Deconvolution2DBuilder<Deconvolution2D, Deconvolution2DBuilderImpl> {
public Deconvolution2D build() {
Deconvolution2D l = new Deconvolution2D(this);
l.initializeConstraints();
return l;
}
}
public static abstract class Deconvolution2DBuilder<C extends Deconvolution2D, B extends Deconvolution2DBuilder<C, B>> extends ConvolutionLayerBuilder<C, B> {
@Override

43
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Deconvolution3D.java
View File

@ -44,7 +44,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class Deconvolution3D extends ConvolutionLayer {
/**
* Set the convolution mode for the Convolution layer. See {@link ConvolutionMode} for more
@ -56,6 +56,15 @@ public class Deconvolution3D extends ConvolutionLayer {
private Convolution3D.DataFormat dataFormat =
Convolution3D.DataFormat.NCDHW; // in libnd4j: 1 - NCDHW, 0 - NDHWC
public static Deconvolution3DBuilder<?, ?> builder() {
return innerBuilder()
.kernelSize(new int[] {2, 2, 2})
.stride(new int[] {1, 1, 1})
.padding(new int[] {0, 0, 0})
.dilation(new int[] {1, 1, 1})
.convolutionDim(3);
}
protected boolean allowCausal() {
// Causal convolution - allowed for 1D only
return false;
@ -134,6 +143,11 @@ public class Deconvolution3D extends ConvolutionLayer {
}
}
// private int[] kernelSize;
// private int[] stride;
// private int[] padding;
// private int[] dilation;
@Override
public InputType getOutputType(int layerIndex, InputType inputType) {
if (inputType == null || inputType.getType() != InputType.Type.CNN3D) {
@ -158,29 +172,16 @@ public class Deconvolution3D extends ConvolutionLayer {
Deconvolution3DLayer.class);
}
//private int[] kernelSize;
//private int[] stride;
//private int[] padding;
//private int[] dilation;
public static abstract class Deconvolution3DBuilder<
public abstract static class Deconvolution3DBuilder<
C extends Deconvolution3D, B extends Deconvolution3DBuilder<C, B>>
extends ConvolutionLayerBuilder<C, B> {
public C build() {
C l = initBuild();
extends ConvolutionLayerBuilder<C, B> {}
private static final class Deconvolution3DBuilderImpl
extends Deconvolution3DBuilder<Deconvolution3D, Deconvolution3DBuilderImpl> {
public Deconvolution3D build() {
Deconvolution3D l = new Deconvolution3D(this);
l.initializeConstraints();
return l;
}
}
public static Deconvolution3DBuilder<?,?> builder() {
return innerBuilder()
.kernelSize(new int[] {2, 2, 2})
.stride(new int[] {1, 1, 1})
.padding(new int[] {0, 0, 0})
.dilation(new int[] {1, 1, 1})
.convolutionDim(3);
}
}

20
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/DepthwiseConvolution2D.java
View File

@ -40,7 +40,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class DepthwiseConvolution2D extends ConvolutionLayer {
/**
* Set channels multiplier for depth-wise convolution
@ -145,21 +145,25 @@ public class DepthwiseConvolution2D extends ConvolutionLayer {
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() {
private static final class DepthwiseConvolution2DBuilderImpl extends DepthwiseConvolution2DBuilder<DepthwiseConvolution2D, DepthwiseConvolution2DBuilderImpl> {
public DepthwiseConvolution2D build() {
DepthwiseConvolution2D l = new DepthwiseConvolution2D(this);
Preconditions.checkState(
depthMultiplier$value > 0,
l.getDepthMultiplier() > 0,
"Depth multiplier must be > 0, got %s",
depthMultiplier$value);
C l = this.initBuild();
l.getDepthMultiplier());
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), l.getPadding());
ConvolutionUtils.validateCnnKernelStridePadding(
l.getKernelSize(), l.getStride(), l.getPadding());
l.initializeConstraints();
return l;
}
}
public abstract static class DepthwiseConvolution2DBuilder<
C extends DepthwiseConvolution2D, B extends DepthwiseConvolution2DBuilder<C, B>>
extends ConvolutionLayerBuilder<C, B> {
@Override
public B kernelSize(int... kernelSize) {

141
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/EmbeddingLayer.java
View File

@ -20,6 +20,8 @@
package org.deeplearning4j.nn.conf.layers;
import java.util.Collection;
import java.util.Map;
import lombok.*;
import lombok.experimental.Accessors;
import lombok.experimental.SuperBuilder;
@ -30,95 +32,47 @@ import org.deeplearning4j.nn.conf.inputs.InputType;
import org.deeplearning4j.nn.conf.memory.LayerMemoryReport;
import org.deeplearning4j.nn.conf.memory.MemoryReport;
import org.deeplearning4j.nn.params.EmbeddingLayerParamInitializer;
import org.deeplearning4j.nn.weights.IWeightInit;
import org.deeplearning4j.nn.weights.WeightInit;
import org.deeplearning4j.nn.weights.embeddings.ArrayEmbeddingInitializer;
import org.deeplearning4j.nn.weights.embeddings.EmbeddingInitializer;
import org.deeplearning4j.nn.weights.embeddings.WeightInitEmbedding;
import org.deeplearning4j.optimize.api.TrainingListener;
import org.nd4j.linalg.activations.Activation;
import org.nd4j.linalg.activations.impl.ActivationIdentity;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import java.util.Collection;
import java.util.Map;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class EmbeddingLayer extends FeedForwardLayer {
/**
* If true: include bias parameters in the layer. False (default): no bias.
*
* @param hasBias If true: include bias parameters in this layer
*/
@Accessors @Builder.Default
private boolean hasBias = false;
@Accessors @Builder.Default private boolean hasBias = false;
/**
*Default to Identity activation - i.e., don't inherit.
* For example, if user sets ReLU as global default, they very likely don't intend to use it for Embedding layer also
*
* Default to Identity activation - i.e., don't inherit. For example, if user sets ReLU as global
* default, they very likely don't intend to use it for Embedding layer also
*/
public static EmbeddingLayerBuilder<?, ?> builder() {
return innerBuilder()
.activation(Activation.IDENTITY);
return innerBuilder().activation(Activation.IDENTITY);
}
public static abstract class EmbeddingLayerBuilder<C extends EmbeddingLayer, B extends EmbeddingLayerBuilder<C,B>>
extends FeedForwardLayerBuilder<C,B>{
public C build() {
C l = initBuild();
l.initializeConstraints();
return l;
}
/**
* Weight initialization scheme to use, for initial weight values
*
* @param weightInit
* @see WeightInit
*/
@Override
public B weightInit(WeightInit weightInit) {
if(weightInit.getWeightInitFunction() instanceof WeightInitEmbedding){
long[] shape = ((WeightInitEmbedding) weightInit.getWeightInitFunction()).shape();
nIn(shape[0]);
nOut(shape[1]);
}
super.weightInit(weightInit);
return self();
}
/**
* Initialize the embedding layer using values from the specified array. Note that the array should have shape
* [vocabSize, vectorSize]. After copying values from the array to initialize the network parameters, the input
* array will be discarded (so that, if necessary, it can be garbage collected)
*
* @param vectors Vectors to initialize the embedding layer with
*/
public B weightInit(INDArray vectors){
weightInit(new ArrayEmbeddingInitializer(vectors));
return self();
}
/**
* Initialize the embedding layer using the specified EmbeddingInitializer - such as a Word2Vec instance
*
* @param embeddingInitializer Source of the embedding layer weights
*/
public B weightInit(EmbeddingInitializer embeddingInitializer) {
var weightIn = new WeightInitEmbedding(embeddingInitializer);
super.weightInit(weightIn);
return self();
}
}
@Override
public Layer instantiate(NeuralNetConfiguration conf, Collection<TrainingListener> trainingListeners,
int layerIndex, INDArray layerParamsView, boolean initializeParams, DataType networkDataType) {
public Layer instantiate(
NeuralNetConfiguration conf,
Collection<TrainingListener> trainingListeners,
int layerIndex,
INDArray layerParamsView,
boolean initializeParams,
DataType networkDataType) {
LayerConfiguration lconf = conf.getFlattenedLayerConfigurations().get(layerIndex);
org.deeplearning4j.nn.layers.feedforward.embedding.EmbeddingLayer ret =
new org.deeplearning4j.nn.layers.feedforward.embedding.EmbeddingLayer(lconf, networkDataType);
new org.deeplearning4j.nn.layers.feedforward.embedding.EmbeddingLayer(
lconf, networkDataType);
runInheritance();
ret.addTrainingListeners(trainingListeners);
@ -149,8 +103,65 @@ public class EmbeddingLayer extends FeedForwardLayer {
// Training: preout op, the only in-place ops on epsilon (from layer above) + assign ops
return new LayerMemoryReport.Builder(name, EmbeddingLayer.class, inputType, outputType)
.standardMemory(numParams, updaterStateSize).workingMemory(0, 0, 0, actElementsPerEx)
.cacheMemory(MemoryReport.CACHE_MODE_ALL_ZEROS, MemoryReport.CACHE_MODE_ALL_ZEROS) //No caching
.standardMemory(numParams, updaterStateSize)
.workingMemory(0, 0, 0, actElementsPerEx)
.cacheMemory(
MemoryReport.CACHE_MODE_ALL_ZEROS, MemoryReport.CACHE_MODE_ALL_ZEROS) // No caching
.build();
}
private static final class EmbeddingLayerBuilderImpl
extends EmbeddingLayerBuilder<EmbeddingLayer, EmbeddingLayerBuilderImpl> {
public EmbeddingLayer build() {
EmbeddingLayer l = new EmbeddingLayer(this);
l.initializeConstraints();
return l;
}
}
public abstract static class EmbeddingLayerBuilder<
C extends EmbeddingLayer, B extends EmbeddingLayerBuilder<C, B>>
extends FeedForwardLayerBuilder<C, B> {
/**
* Weight initialization scheme to use, for initial weight values
*
* @param weightInit
* @see WeightInit
*/
@Override
public B weightInit(WeightInit weightInit) {
if (weightInit.getWeightInitFunction() instanceof WeightInitEmbedding) {
long[] shape = ((WeightInitEmbedding) weightInit.getWeightInitFunction()).shape();
nIn(shape[0]);
nOut(shape[1]);
}
super.weightInit(weightInit);
return self();
}
/**
* Initialize the embedding layer using values from the specified array. Note that the array
* should have shape [vocabSize, vectorSize]. After copying values from the array to initialize
* the network parameters, the input array will be discarded (so that, if necessary, it can be
* garbage collected)
*
* @param vectors Vectors to initialize the embedding layer with
*/
public B weightInit(INDArray vectors) {
weightInit(new ArrayEmbeddingInitializer(vectors));
return self();
}
/**
* Initialize the embedding layer using the specified EmbeddingInitializer - such as a Word2Vec
* instance
*
* @param embeddingInitializer Source of the embedding layer weights
*/
public B weightInit(EmbeddingInitializer embeddingInitializer) {
var weightIn = new WeightInitEmbedding(embeddingInitializer);
super.weightInit(weightIn);
return self();
}
}
}

13
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/EmbeddingSequenceLayer.java
View File

@ -46,7 +46,7 @@ import java.util.Map;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class EmbeddingSequenceLayer extends FeedForwardLayer {
/**
* Set input sequence length for this embedding layer.
@ -70,13 +70,16 @@ public class EmbeddingSequenceLayer extends FeedForwardLayer {
@Builder.Default private boolean inferInputLength = false; // use input length as provided by input data
@Builder.Default private RNNFormat outputDataFormat = RNNFormat.NCW; //Default value for older deserialized models
public static abstract class EmbeddingSequenceLayerBuilder<C extends EmbeddingSequenceLayer, B extends EmbeddingSequenceLayerBuilder<C, B>>
extends FeedForwardLayerBuilder<C, B> {
public C build() {
C l = initBuild();
private static final class EmbeddingSequenceLayerBuilderImpl extends EmbeddingSequenceLayerBuilder<EmbeddingSequenceLayer, EmbeddingSequenceLayerBuilderImpl> {
public EmbeddingSequenceLayer build() {
EmbeddingSequenceLayer l = new EmbeddingSequenceLayer(this);
l.initializeConstraints();
return l;
}
}
public static abstract class EmbeddingSequenceLayerBuilder<C extends EmbeddingSequenceLayer, B extends EmbeddingSequenceLayerBuilder<C, B>>
extends FeedForwardLayerBuilder<C, B> {
public B weightInit(IWeightInit weightInit){
if(weightInit instanceof WeightInitEmbedding){

43
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/GravesBidirectionalLSTM.java
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;
@ -36,23 +37,21 @@ import org.nd4j.linalg.activations.impl.ActivationSigmoid;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import java.util.*;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@Deprecated
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class GravesBidirectionalLSTM extends BaseRecurrentLayer {
public static abstract class GravesBidirectionalLSTMBuilder<C extends GravesBidirectionalLSTM, B extends
GravesBidirectionalLSTMBuilder<C, B>> extends BaseRecurrentLayerBuilder<C, B> {
public C build() {
C l = this.initBuild();
l.initializeConstraints();
return l;
}
}
/**
* 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 GravesBidirectionalLSTM will be used
*
*/
@Builder.Default
protected boolean helperAllowFallback = true;
/**
* Set forget gate bias initalizations. Values in range 1-5 can potentially help with learning or longer-term
* dependencies.
@ -66,15 +65,6 @@ public class GravesBidirectionalLSTM extends BaseRecurrentLayer {
*/
@Builder.Default
private IActivation gateActivationFunction = new ActivationSigmoid();
/**
* 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 GravesBidirectionalLSTM will be used
*
*/
@Builder.Default
protected boolean helperAllowFallback = true;
@Override
protected void initializeConstraints() {
@ -121,5 +111,18 @@ public class GravesBidirectionalLSTM extends BaseRecurrentLayer {
return LSTMHelpers.getMemoryReport(this, inputType);
}
private static final class GravesBidirectionalLSTMBuilderImpl extends GravesBidirectionalLSTMBuilder<GravesBidirectionalLSTM, GravesBidirectionalLSTMBuilderImpl> {
public GravesBidirectionalLSTM build() {
GravesBidirectionalLSTM l = new GravesBidirectionalLSTM(this);
l.initializeConstraints();
return l;
}
}
public static abstract class GravesBidirectionalLSTMBuilder<C extends GravesBidirectionalLSTM, B extends
GravesBidirectionalLSTMBuilder<C, B>> extends BaseRecurrentLayerBuilder<C, B> {
}
}

11
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/GravesLSTM.java
View File

@ -43,7 +43,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
@Deprecated
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class GravesLSTM extends AbstractLSTM {
private double forgetGateBiasInit;
@ -103,9 +103,12 @@ public class GravesLSTM extends AbstractLSTM {
public abstract static class GravesLSTMBuilder<
C extends GravesLSTM, B extends GravesLSTMBuilder<C, B>>
extends AbstractLSTMBuilder<C, B> {
public C build() {
C l = initBuild();
extends AbstractLSTMBuilder<C, B> {}
private static final class GravesLSTMBuilderImpl
extends GravesLSTMBuilder<GravesLSTM, GravesLSTMBuilderImpl> {
public GravesLSTM build() {
GravesLSTM l = new GravesLSTM(this);
l.initializeConstraints();
return l;
}

44
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/LSTM.java
View File

@ -20,6 +20,10 @@
package org.deeplearning4j.nn.conf.layers;
import java.util.ArrayList;
import java.util.Collection;
import java.util.Collections;
import java.util.Map;
import lombok.*;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.Layer;
@ -31,31 +35,17 @@ import org.deeplearning4j.nn.conf.memory.LayerMemoryReport;
import org.deeplearning4j.nn.layers.recurrent.LSTMHelpers;
import org.deeplearning4j.nn.params.LSTMParamInitializer;
import org.deeplearning4j.optimize.api.TrainingListener;
import org.nd4j.linalg.activations.Activation;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import java.util.ArrayList;
import java.util.Collection;
import java.util.Collections;
import java.util.Map;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class LSTM extends AbstractLSTM {
private double forgetGateBiasInit;
public static abstract class LSTMBuilder<C extends LSTM, B extends LSTMBuilder<C, B>> extends AbstractLSTMBuilder<C, B> {
@Override public C build() {
C l = this.initBuild();
l.initializeConstraints();
return l;
}
}
@Override
protected void initializeConstraints() {
super.initializeConstraints();
@ -72,13 +62,19 @@ public class LSTM extends AbstractLSTM {
}
@Override
public Layer instantiate(NeuralNetConfiguration conf, Collection<TrainingListener> trainingListeners,
int layerIndex, INDArray layerParamsView, boolean initializeParams, DataType networkDataType) {
public Layer instantiate(
NeuralNetConfiguration conf,
Collection<TrainingListener> trainingListeners,
int layerIndex,
INDArray layerParamsView,
boolean initializeParams,
DataType networkDataType) {
LayerValidation.assertNInNOutSet("LSTM", getName(), layerIndex, getNIn(), getNOut());
LayerConfiguration lconf = conf.getFlattenedLayerConfigurations().get(layerIndex);
runInheritance();
org.deeplearning4j.nn.layers.recurrent.LSTM ret = new org.deeplearning4j.nn.layers.recurrent.LSTM(lconf, networkDataType);
org.deeplearning4j.nn.layers.recurrent.LSTM ret =
new org.deeplearning4j.nn.layers.recurrent.LSTM(lconf, networkDataType);
ret.addTrainingListeners(trainingListeners);
ret.setIndex(layerIndex);
ret.setParamsViewArray(layerParamsView);
@ -98,4 +94,16 @@ public class LSTM extends AbstractLSTM {
// TODO - CuDNN etc
return LSTMHelpers.getMemoryReport(this, inputType);
}
public abstract static class LSTMBuilder<C extends LSTM, B extends LSTMBuilder<C, B>>
extends AbstractLSTMBuilder<C, B> {}
private static final class LSTMBuilderImpl extends LSTMBuilder<LSTM, LSTMBuilderImpl> {
@Override
public LSTM build() {
LSTM l = new LSTM(this);
l.initializeConstraints();
return l;
}
}
}

10
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/LayerConfiguration.java
View File

@ -354,15 +354,5 @@ public abstract class LayerConfiguration
biasConstraints = Arrays.asList(constraints);
return self();
}
/**
* we are doing this to avoid BUG https://github.com/projectlombok/lombok/issues/3419 as some
* child classes may specify their own buildMethodName in @SuperBuilder, but we use only
* "initBuild" here consequently
* @return
*/
public C initBuild() {
return build();
}
}
}

31
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/LearnedSelfAttentionLayer.java
View File

@ -41,7 +41,7 @@ import org.nd4j.linalg.factory.Nd4j;
@Data
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class LearnedSelfAttentionLayer extends SameDiffLayer {
private static final String WEIGHT_KEY_QUERY_PROJECTION = "Wq";
private static final String WEIGHT_KEY_KEY_PROJECTION = "Wk";
@ -173,19 +173,24 @@ public class LearnedSelfAttentionLayer extends SameDiffLayer {
public static abstract class LearnedSelfAttentionLayerBuilder<
C extends LearnedSelfAttentionLayer, B extends LearnedSelfAttentionLayerBuilder<C, B>>
extends SameDiffLayerBuilder<C, B> {
public C build() {
Preconditions.checkArgument(
this.projectInput || this.nHeads == 1, "projectInput must be true when nHeads != 1");
Preconditions.checkArgument(
this.projectInput || nIn == nOut, "nIn must be equal to nOut when projectInput is false");
Preconditions.checkArgument(
!this.projectInput || nOut != 0, "nOut must be specified when projectInput is true");
Preconditions.checkArgument(
this.nOut % nHeads == 0 || headSize > 0,
"nOut isn't divided by nHeads cleanly. Specify the headSize manually.");
Preconditions.checkArgument(this.nQueries > 0, "You must set numQueries.");
return initBuild();
}
private static final class LearnedSelfAttentionLayerBuilderImpl extends LearnedSelfAttentionLayerBuilder<LearnedSelfAttentionLayer, LearnedSelfAttentionLayerBuilderImpl> {
public LearnedSelfAttentionLayer build() {
LearnedSelfAttentionLayer l = new LearnedSelfAttentionLayer(this);
Preconditions.checkArgument(
l.isProjectInput() || l.getNHeads() == 1, "projectInput must be true when nHeads != 1");
Preconditions.checkArgument(
l.isProjectInput() || l.getNIn() == l.getNOut(), "nIn must be equal to nOut when projectInput is false");
Preconditions.checkArgument(
!l.isProjectInput() || l.getNOut() != 0, "nOut must be specified when projectInput is true");
Preconditions.checkArgument(
l.getNOut() % l.getNHeads() == 0 || l.getHeadSize() > 0,
"nOut isn't divided by nHeads cleanly. Specify the headSize manually.");
Preconditions.checkArgument(l.getNQueries() > 0, "You must set numQueries.");
return l;
}
}
}

30
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/LocallyConnected1D.java
View File

@ -48,19 +48,9 @@ import org.nd4j.linalg.factory.Nd4j;
@Data
@EqualsAndHashCode(callSuper = true)
@JsonIgnoreProperties({"paramShapes"})
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class LocallyConnected1D extends SameDiffLayer {
public static abstract class LocallyConnected1DBuilder<C extends LocallyConnected1D, B extends LocallyConnected1DBuilder<C, B>> extends
SameDiffLayerBuilder<C, B> {
public C build() {
Convolution1DUtils.validateConvolutionModePadding(convolutionMode$value, padding$value);
Convolution1DUtils.validateCnn1DKernelStridePadding(kernelSize$value, stride$value, padding$value);
C l = initBuild();
return l;
}
}
private static final List<String> WEIGHT_KEYS =
Collections.singletonList(ConvolutionParamInitializer.WEIGHT_KEY);
private static final List<String> BIAS_KEYS =
@ -89,10 +79,8 @@ public class LocallyConnected1D extends SameDiffLayer {
private int paddingR; // Right/bottom padding
/** Convolution mode for the layer. See {@link ConvolutionMode} for details */
@Builder.Default private ConvolutionMode convolutionMode = ConvolutionMode.Same;
/** Dilation for the layer */
@Builder.Default private int dilation = 1;
/** If true (default is false) the layer will have a bias */
@Builder.Default private boolean hasBias = true;
@ -272,4 +260,20 @@ public class LocallyConnected1D extends SameDiffLayer {
convolutionMode = global_conf.getConvolutionMode();
}
}
private static final class LocallyConnected1DBuilderImpl
extends LocallyConnected1DBuilder<LocallyConnected1D, LocallyConnected1DBuilderImpl> {
public LocallyConnected1D build() {
LocallyConnected1D l = new LocallyConnected1D(this);
Convolution1DUtils.validateConvolutionModePadding(l.getConvolutionMode(), l.getPadding());
Convolution1DUtils.validateCnn1DKernelStridePadding(
l.getKernelSize(), l.getStride(), l.getPadding());
return l;
}
}
public abstract static class LocallyConnected1DBuilder<
C extends LocallyConnected1D, B extends LocallyConnected1DBuilder<C, B>>
extends SameDiffLayerBuilder<C, B> {}
}

19
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/LocallyConnected2D.java
View File

@ -41,7 +41,6 @@ import org.nd4j.autodiff.samediff.SDVariable;
import org.nd4j.autodiff.samediff.SameDiff;
import org.nd4j.enums.PadMode;
import org.nd4j.linalg.activations.Activation;
import org.nd4j.linalg.activations.IActivation;
import org.nd4j.linalg.api.memory.MemoryWorkspace;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.factory.Nd4j;
@ -49,7 +48,7 @@ import org.nd4j.linalg.factory.Nd4j;
@Data
@EqualsAndHashCode(callSuper = true)
@JsonIgnoreProperties({"paramShapes"})
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class LocallyConnected2D extends SameDiffLayer {
private static final List<String> WEIGHT_KEYS =
@ -318,14 +317,18 @@ public class LocallyConnected2D extends SameDiffLayer {
}
}
public static abstract class LocallyConnected2DBuilder<
C extends LocallyConnected2D, B extends LocallyConnected2DBuilder<C, B>>
extends SameDiffLayerBuilder<C, B> {
public C build() {
featureDim(kernel$value[0] * kernel$value[1] * (int) nIn);
C l = initBuild();
private static final class LocallyConnected2DBuilderImpl
extends LocallyConnected2DBuilder<LocallyConnected2D, LocallyConnected2DBuilderImpl> {
public LocallyConnected2D build() {
LocallyConnected2D l = new LocallyConnected2D(this);
l.setFeatureDim(l.getKernel()[0] * l.getKernel()[1] * (int) l.getNIn());
return l;
}
}
public abstract static class LocallyConnected2DBuilder<
C extends LocallyConnected2D, B extends LocallyConnected2DBuilder<C, B>>
extends SameDiffLayerBuilder<C, B> {
public B kernelSize(int... kernel) {
this.kernel$value = ValidationUtils.validate2NonNegative(kernel, false, "kernel");

17
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/OutputLayer.java
View File

@ -22,7 +22,6 @@ package org.deeplearning4j.nn.conf.layers;
import java.util.Collection;
import java.util.Map;
import lombok.*;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.Layer;
@ -33,13 +32,12 @@ import org.deeplearning4j.optimize.api.TrainingListener;
import org.nd4j.linalg.activations.Activation;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.lossfunctions.ILossFunction;
import org.nd4j.linalg.lossfunctions.LossFunctions;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class OutputLayer extends BaseOutputLayer {
{ // Set default activation function to softmax (to match default loss function MCXENT)
@ -82,15 +80,16 @@ public class OutputLayer extends BaseOutputLayer {
return DefaultParamInitializer.getInstance();
}
public static abstract class OutputLayerBuilder<
public abstract static class OutputLayerBuilder<
C extends OutputLayer, B extends OutputLayerBuilder<C, B>>
extends BaseOutputLayerBuilder<C, B> {
public C build() {
C l = this.initBuild();
extends BaseOutputLayerBuilder<C, B> {}
private static final class OutputLayerBuilderImpl
extends OutputLayerBuilder<OutputLayer, OutputLayerBuilderImpl> {
public OutputLayer build() {
OutputLayer l = new OutputLayer(this);
l.initializeConstraints();
return l;
}
}
}

15
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/PReLULayer.java
View File

@ -40,7 +40,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class PReLULayer extends BaseLayerConfiguration {
/**
* Explicitly set input shape of incoming activations so that parameters can be initialized
@ -129,14 +129,17 @@ public class PReLULayer extends BaseLayerConfiguration {
.build();
}
public static abstract class PReLULayerBuilder<
C extends PReLULayer, B extends PReLULayerBuilder<C, B>>
extends BaseLayerConfigurationBuilder<C, B> {
public C build() {
C l = initBuild();
private static final class PReLULayerBuilderImpl extends PReLULayerBuilder<PReLULayer, PReLULayerBuilderImpl> {
public PReLULayer build() {
PReLULayer l = new PReLULayer(this);
l.initializeConstraints();
return l;
}
}
public static abstract class PReLULayerBuilder<
C extends PReLULayer, B extends PReLULayerBuilder<C, B>>
extends BaseLayerConfigurationBuilder<C, B> {
/**
* Explicitly set input shape of incoming activations so that parameters can be initialized

2
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Pooling1D.java
View File

@ -35,6 +35,6 @@ import lombok.experimental.SuperBuilder;
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class Pooling1D extends Subsampling1DLayer {
}

2
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Pooling2D.java
View File

@ -35,6 +35,6 @@ import lombok.experimental.SuperBuilder;
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class Pooling2D extends SubsamplingLayer {
}

21
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/PrimaryCapsules.java
View File

@ -41,7 +41,7 @@ import org.nd4j.linalg.factory.Nd4j;
@Data
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class PrimaryCapsules extends SameDiffLayer {
private static final String WEIGHT_PARAM = "weight";
@ -335,7 +335,7 @@ public class PrimaryCapsules extends SameDiffLayer {
}
}
public static abstract class PrimaryCapsulesBuilder<
public abstract static class PrimaryCapsulesBuilder<
C extends PrimaryCapsules, B extends PrimaryCapsulesBuilder<C, B>>
extends SameDiffLayerBuilder<C, B> {
@ -396,27 +396,30 @@ public class PrimaryCapsules extends SameDiffLayer {
this.useLeakyReLU$set = true;
return self();
}
}
public C build() {
C l = initBuild();
if (capsuleDimensions <= 0 || channels$value <= 0) {
private static final class PrimaryCapsulesBuilderImpl
extends PrimaryCapsulesBuilder<PrimaryCapsules, PrimaryCapsulesBuilderImpl> {
public PrimaryCapsules build() {
PrimaryCapsules l = new PrimaryCapsules(this);
if (l.getCapsuleDimensions() <= 0 || l.getChannels() <= 0) {
throw new IllegalArgumentException(
"Invalid configuration for Primary Capsules (layer name = \""
+ l.getName()
+ "\"):"
+ " capsuleDimensions and channels must be > 0. Got: "
+ capsuleDimensions
+ l.getCapsuleDimensions()
+ ", "
+ channels$value);
+ l.getChannels());
}
if (capsules < 0) {
if (l.getCapsules() < 0) {
throw new IllegalArgumentException(
"Invalid configuration for Capsule ILayer (layer name = \""
+ l.getName()
+ "\"):"
+ " capsules must be >= 0 if set. Got: "
+ capsules);
+ l.getCapsules());
}
return l;
}

22
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/RecurrentAttentionLayer.java
View File

@ -43,21 +43,25 @@ import java.util.Map;
@Data
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class RecurrentAttentionLayer extends SameDiffLayer {
private static final class RecurrentAttentionLayerBuilderImpl extends RecurrentAttentionLayerBuilder<RecurrentAttentionLayer, RecurrentAttentionLayerBuilderImpl> {
public RecurrentAttentionLayer build() {
RecurrentAttentionLayer l = new RecurrentAttentionLayer(this);
Preconditions.checkArgument(l.isProjectInput() || l.getNHeads() == 1, "projectInput must be true when nHeads != 1");
Preconditions.checkArgument(l.isProjectInput() || l.getNIn() == l.getNOut(), "nIn must be equal to nOut when projectInput is false");
Preconditions.checkArgument(!l.isProjectInput() || l.getNOut() != 0, "nOut must be specified when projectInput is true");
Preconditions.checkArgument(l.getNOut() % l.getNHeads() == 0 || l.getNHeads() > 0, "nOut isn't divided by nHeads cleanly. Specify the headSize manually.");
return l;
}
}
public static abstract class RecurrentAttentionLayerBuilder<C extends RecurrentAttentionLayer, B extends RecurrentAttentionLayerBuilder<C,B>>
extends SameDiffLayerBuilder<C,B> {
public C build() {
Preconditions.checkArgument(this.projectInput$value || this.nHeads == 1, "projectInput must be true when nHeads != 1");
Preconditions.checkArgument(this.projectInput$value || nIn == nOut, "nIn must be equal to nOut when projectInput is false");
Preconditions.checkArgument(!this.projectInput$value || nOut != 0, "nOut must be specified when projectInput is true");
Preconditions.checkArgument(this.nOut % nHeads == 0 || headSize > 0, "nOut isn't divided by nHeads cleanly. Specify the headSize manually.");
C l = initBuild();
return l;
}
}
/**

47
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/RnnOutputLayer.java
View File

@ -24,7 +24,6 @@ import java.util.Collection;
import java.util.Map;
import lombok.Data;
import lombok.EqualsAndHashCode;
import lombok.NoArgsConstructor;
import lombok.ToString;
import lombok.experimental.SuperBuilder;
import org.deeplearning4j.nn.api.Layer;
@ -42,12 +41,12 @@ import org.nd4j.linalg.lossfunctions.LossFunctions;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class RnnOutputLayer extends BaseOutputLayer {
/**
* @param rnnDataFormat Data format expected by the layer. NCW = [miniBatchSize, size, timeSeriesLength],
* NWC = [miniBatchSize, timeSeriesLength, size]. Defaults to NCW.
* @param rnnDataFormat Data format expected by the layer. NCW = [miniBatchSize, size,
* timeSeriesLength], NWC = [miniBatchSize, timeSeriesLength, size]. Defaults to NCW.
*/
private RNNFormat dataFormat;
@ -59,13 +58,17 @@ public class RnnOutputLayer extends BaseOutputLayer {
* @param lossFn Loss function for the output layer
*/
public static RnnOutputLayerBuilder<?, ?> builder(LossFunctions.LossFunction lossFn) {
return innerBuilder()
.lossFunction(lossFn);
return innerBuilder().lossFunction(lossFn);
}
@Override
public Layer instantiate(NeuralNetConfiguration conf, Collection<TrainingListener> trainingListeners,
int layerIndex, INDArray layerParamsView, boolean initializeParams, DataType networkDataType) {
public Layer instantiate(
NeuralNetConfiguration conf,
Collection<TrainingListener> trainingListeners,
int layerIndex,
INDArray layerParamsView,
boolean initializeParams,
DataType networkDataType) {
LayerValidation.assertNInNOutSet("RnnOutputLayer", getName(), layerIndex, getNIn(), getNOut());
LayerConfiguration lconf = conf.getFlattenedLayerConfigurations().get(layerIndex);
@ -88,8 +91,13 @@ public class RnnOutputLayer extends BaseOutputLayer {
@Override
public InputType getOutputType(int layerIndex, InputType inputType) {
if (inputType == null || inputType.getType() != InputType.Type.RNN) {
throw new IllegalStateException("Invalid input type for RnnOutputLayer (layer index = " + layerIndex
+ ", layer name=\"" + getName() + "\"): Expected RNN input, got " + inputType);
throw new IllegalStateException(
"Invalid input type for RnnOutputLayer (layer index = "
+ layerIndex
+ ", layer name=\""
+ getName()
+ "\"): Expected RNN input, got "
+ inputType);
}
InputType.InputTypeRecurrent itr = (InputType.InputTypeRecurrent) inputType;
@ -99,8 +107,11 @@ public class RnnOutputLayer extends BaseOutputLayer {
@Override
public void setNIn(InputType inputType, boolean override) {
if (inputType == null || inputType.getType() != InputType.Type.RNN) {
throw new IllegalStateException("Invalid input type for RnnOutputLayer (layer name=\"" + getName()
+ "\"): Expected RNN input, got " + inputType);
throw new IllegalStateException(
"Invalid input type for RnnOutputLayer (layer name=\""
+ getName()
+ "\"): Expected RNN input, got "
+ inputType);
}
InputType.InputTypeRecurrent r = (InputType.InputTypeRecurrent) inputType;
@ -118,12 +129,16 @@ public class RnnOutputLayer extends BaseOutputLayer {
return InputTypeUtil.getPreprocessorForInputTypeRnnLayers(inputType, dataFormat, getName());
}
public static abstract class RnnOutputLayerBuilder<C extends RnnOutputLayer, B extends RnnOutputLayerBuilder<C, B>> extends BaseOutputLayerBuilder<C, B> {
public C build() {
C l = this.initBuild();
public abstract static class RnnOutputLayerBuilder<
C extends RnnOutputLayer, B extends RnnOutputLayerBuilder<C, B>>
extends BaseOutputLayerBuilder<C, B> {}
private static final class RnnOutputLayerBuilderImpl
extends RnnOutputLayerBuilder<RnnOutputLayer, RnnOutputLayerBuilderImpl> {
public RnnOutputLayer build() {
RnnOutputLayer l = new RnnOutputLayer(this);
l.initializeConstraints();
return l;
}
}
}

2
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/SelfAttentionLayer.java
View File

@ -38,7 +38,7 @@ import org.nd4j.linalg.factory.Nd4j;
@Data
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class SelfAttentionLayer extends SameDiffLayer {
private static final String WEIGHT_KEY_QUERY_PROJECTION = "Wq";
private static final String WEIGHT_KEY_KEY_PROJECTION = "Wk";

2
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/SeparableConvolution2D.java
View File

@ -44,7 +44,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class SeparableConvolution2D extends ConvolutionLayer {
/**
* Set constraints to be applied to the point-wise convolution weight parameters of this layer.

15
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Subsampling1DLayer.java
View File

@ -50,7 +50,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class Subsampling1DLayer extends SubsamplingLayer {
@Override
@ -153,11 +153,9 @@ public class Subsampling1DLayer extends SubsamplingLayer {
return true;
}
public static abstract class Subsampling1DLayerBuilder<C extends Subsampling1DLayer, B extends Subsampling1DLayerBuilder<C, B>> extends
SubsamplingLayerBuilder<C, B> {
public C build() {
C l = this.initBuild();
private static final class Subsampling1DLayerBuilderImpl extends Subsampling1DLayerBuilder<Subsampling1DLayer, Subsampling1DLayerBuilderImpl> {
public Subsampling1DLayer build() {
Subsampling1DLayer l =new Subsampling1DLayer(this);
if (l.getPoolingType() == org.deeplearning4j.nn.conf.layers.PoolingType.PNORM && l.getPnorm() <= 0) {
throw new IllegalStateException(
"Incorrect Subsampling config: p-norm must be set when using PoolingType.PNORM");
@ -167,6 +165,11 @@ public class Subsampling1DLayer extends SubsamplingLayer {
return l;
}
}
public static abstract class Subsampling1DLayerBuilder<C extends Subsampling1DLayer, B extends Subsampling1DLayerBuilder<C, B>> extends
SubsamplingLayerBuilder<C, B> {
/**
*
* @param kernelSize

19
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Subsampling3DLayer.java
View File

@ -45,7 +45,7 @@ import org.nd4j.linalg.learning.regularization.Regularization;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(builderMethodName = "innerBuilder", buildMethodName = "initBuild")
@SuperBuilder(builderMethodName = "innerBuilder")
public class Subsampling3DLayer extends NoParamLayer {
@Builder.Default protected ConvolutionMode convolutionMode = ConvolutionMode.Truncate;
@ -304,17 +304,22 @@ public class Subsampling3DLayer extends NoParamLayer {
return self();
}
public C build() {
if (kernelSize.length != 3) {
}
private static final class Subsampling3DLayerBuilderImpl extends Subsampling3DLayerBuilder<Subsampling3DLayer, Subsampling3DLayerBuilderImpl> {
public Subsampling3DLayer build() {
Subsampling3DLayer l = new Subsampling3DLayer(this);
if (l.getKernelSize().length != 3) {
throw new IllegalArgumentException("Kernel size must be length 3");
}
if (stride.length != 3) {
if (l.getStride().length != 3) {
throw new IllegalArgumentException("Invalid stride, must be length 3");
}
C l = this.initBuild();
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), padding);
Convolution3DUtils.validateCnn3DKernelStridePadding(kernelSize, stride, padding);
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), l.getPadding());
Convolution3DUtils.validateCnn3DKernelStridePadding(l.getKernelSize(), l.getStride(), l.getPadding());
return l;
}
}

41
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/SubsamplingLayer.java
View File

@ -45,7 +45,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
@Data
@ToString(callSuper = true)
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(buildMethodName = "initBuild", builderMethodName = "innerBuilder")
@SuperBuilder(builderMethodName = "innerBuilder")
public class SubsamplingLayer extends NoParamLayer {
public static final CNN2DFormat DEFAULT_FORMAT = CNN2DFormat.NCHW;
@ -425,25 +425,7 @@ public class SubsamplingLayer extends NoParamLayer {
return self();
}
public C build() {
if (kernelSize$value.length != 2) {
throw new IllegalArgumentException("Kernel size of should be rows x columns (a 2d array)");
}
if (stride$value.length != 2) {
throw new IllegalArgumentException("Invalid stride, must be length 2");
}
if (poolingType$value == org.deeplearning4j.nn.conf.layers.PoolingType.PNORM && pnorm <= 0) {
throw new IllegalStateException(
"Incorrect Subsampling config: p-norm must be set when using PoolingType.PNORM");
}
ConvolutionUtils.validateConvolutionModePadding(convolutionMode$value, padding$value);
ConvolutionUtils.validateCnnKernelStridePadding(
kernelSize$value, stride$value, padding$value);
C l = initBuild();
return l;
}
public B setConvolutionMode(ConvolutionMode convolutionMode){
Preconditions.checkState(allowCausal$value || convolutionMode$value != ConvolutionMode.Causal, "Causal convolution mode can only be used with 1D" +
@ -459,4 +441,25 @@ public class SubsamplingLayer extends NoParamLayer {
return self();
}
}
private static final class SubsamplingLayerBuilderImpl extends SubsamplingLayerBuilder<SubsamplingLayer, SubsamplingLayerBuilderImpl> {
public SubsamplingLayer build() {
SubsamplingLayer l = new SubsamplingLayer(this);
if (l.getKernelSize().length != 2) {
throw new IllegalArgumentException("Kernel size of should be rows x columns (a 2d array)");
}
if (l.getStride().length != 2) {
throw new IllegalArgumentException("Invalid stride, must be length 2");
}
if (l.getPoolingType() == org.deeplearning4j.nn.conf.layers.PoolingType.PNORM && l.getPnorm() <= 0) {
throw new IllegalStateException(
"Incorrect Subsampling config: p-norm must be set when using PoolingType.PNORM");
}
ConvolutionUtils.validateConvolutionModePadding(l.getConvolutionMode(), l.getPadding());
ConvolutionUtils.validateCnnKernelStridePadding(
l.getKernelSize(), l.getStride(), l.getPadding());
return l;
}
}
}

2
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/ZeroPaddingLayer.java
View File

@ -41,7 +41,7 @@ import java.util.Map;
@Data
@EqualsAndHashCode(callSuper = true)
@SuperBuilder(builderMethodName = "innerBuilder", buildMethodName = "initBuild")
@SuperBuilder(builderMethodName = "innerBuilder")
public class ZeroPaddingLayer extends NoParamLayer {
/**
* @param padding Padding value for top, bottom, left, and right. Must be length 4 array

2
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/objdetect/Yolo2OutputLayer.java
View File

@ -48,7 +48,7 @@ import org.nd4j.serde.jackson.shaded.NDArrayTextSerializer;
@EqualsAndHashCode(callSuper = false)
@SuperBuilder(buildMethodName = "initBuild")
@SuperBuilder
public class Yolo2OutputLayer extends LayerConfiguration {
/**

3
cavis-dnn/cavis-dnn-nn/src/main/java/org/deeplearning4j/nn/conf/layers/wrapper/BaseWrapperLayerConfiguration.java
View File

@ -43,8 +43,7 @@ import org.nd4j.linalg.learning.regularization.Regularization;
public abstract class BaseWrapperLayerConfiguration extends LayerConfiguration {
/** The configuration to of another layer to wrap */
@Getter @Setter
protected LayerConfiguration underlying;
@Getter @Setter protected LayerConfiguration underlying;
/**
* Set the net configuration for this configuration as well as for the underlying layer (if not