org.deeplearning4j.zoo.model.Xception Maven / Gradle / Ivy
package org.deeplearning4j.zoo.model;
import lombok.AllArgsConstructor;
import lombok.Builder;
import lombok.NoArgsConstructor;
import org.deeplearning4j.nn.api.Model;
import org.deeplearning4j.nn.api.OptimizationAlgorithm;
import org.deeplearning4j.nn.conf.*;
import org.deeplearning4j.nn.conf.distribution.NormalDistribution;
import org.deeplearning4j.nn.conf.graph.ElementWiseVertex;
import org.deeplearning4j.nn.conf.inputs.InputType;
import org.deeplearning4j.nn.conf.layers.*;
import org.deeplearning4j.nn.graph.ComputationGraph;
import org.deeplearning4j.nn.weights.WeightInit;
import org.deeplearning4j.zoo.ModelMetaData;
import org.deeplearning4j.zoo.PretrainedType;
import org.deeplearning4j.zoo.ZooModel;
import org.deeplearning4j.zoo.ZooType;
import org.nd4j.linalg.activations.Activation;
import org.nd4j.linalg.learning.config.AdaDelta;
import org.nd4j.linalg.learning.config.AdaGrad;
import org.nd4j.linalg.learning.config.IUpdater;
import org.nd4j.linalg.lossfunctions.LossFunctions;
/**
* U-Net
*
* An implementation of Xception in Deeplearning4j. A novel deep convolutional neural network architecture inspired by Inception, where Inception modules have been replaced with depthwise separable convolutions.
*
* Paper: https://arxiv.org/abs/1610.02357
* ImageNet weights for this model are available and have been converted from https://keras.io/applications/.
*
* @author Justin Long (crockpotveggies)
*
*/
@AllArgsConstructor
@Builder
public class Xception extends ZooModel {
@Builder.Default private long seed = 1234;
@Builder.Default private int[] inputShape = new int[] {3, 299, 299};
@Builder.Default private int numClasses = 0;
@Builder.Default private WeightInit weightInit = WeightInit.RELU;
@Builder.Default private IUpdater updater = new AdaDelta();
@Builder.Default private CacheMode cacheMode = CacheMode.NONE;
@Builder.Default private WorkspaceMode workspaceMode = WorkspaceMode.ENABLED;
@Builder.Default private ConvolutionLayer.AlgoMode cudnnAlgoMode = ConvolutionLayer.AlgoMode.PREFER_FASTEST;
private Xception() {}
@Override
public String pretrainedUrl(PretrainedType pretrainedType) {
if (pretrainedType == PretrainedType.IMAGENET)
return "http://blob.deeplearning4j.org/models/xception_dl4j_inference.v2.zip";
else
return null;
}
@Override
public long pretrainedChecksum(PretrainedType pretrainedType) {
if (pretrainedType == PretrainedType.IMAGENET)
return 3277876097L;
else
return 0L;
}
@Override
public Class modelType() {
return ComputationGraph.class;
}
@Override
public ComputationGraph init() {
ComputationGraphConfiguration.GraphBuilder graph = graphBuilder();
graph.addInputs("input").setInputTypes(InputType.convolutional(inputShape[2], inputShape[1], inputShape[0]));
ComputationGraphConfiguration conf = graph.build();
ComputationGraph model = new ComputationGraph(conf);
model.init();
return model;
}
public ComputationGraphConfiguration.GraphBuilder graphBuilder() {
ComputationGraphConfiguration.GraphBuilder graph = new NeuralNetConfiguration.Builder().seed(seed)
.optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
.updater(updater)
.weightInit(weightInit)
.l2(4e-5)
.miniBatch(true)
.cacheMode(cacheMode)
.trainingWorkspaceMode(workspaceMode)
.inferenceWorkspaceMode(workspaceMode)
.convolutionMode(ConvolutionMode.Truncate)
.graphBuilder();
graph
// block1
.addLayer("block1_conv1", new ConvolutionLayer.Builder(3,3).stride(2,2).nOut(32).hasBias(false)
.cudnnAlgoMode(cudnnAlgoMode).build(), "input")
.addLayer("block1_conv1_bn", new BatchNormalization(), "block1_conv1")
.addLayer("block1_conv1_act", new ActivationLayer(Activation.RELU), "block1_conv1_bn")
.addLayer("block1_conv2", new ConvolutionLayer.Builder(3,3).stride(1,1).nOut(64).hasBias(false)
.cudnnAlgoMode(cudnnAlgoMode).build(), "block1_conv1_act")
.addLayer("block1_conv2_bn", new BatchNormalization(), "block1_conv2")
.addLayer("block1_conv2_act", new ActivationLayer(Activation.RELU), "block1_conv1_bn")
// residual1
.addLayer("residual1_conv", new ConvolutionLayer.Builder(1,1).stride(2,2).nOut(128).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "block1_conv2_act")
.addLayer("residual1", new BatchNormalization(), "residual1_conv")
// block2
.addLayer("block2_sepconv1", new SeparableConvolution2D.Builder(3,3).nOut(128).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "block1_conv1_bn")
.addLayer("block2_sepconv1_bn", new BatchNormalization(), "block2_sepconv1")
.addLayer("block2_sepconv1_act",new ActivationLayer(Activation.RELU), "block2_sepconv1_bn")
.addLayer("block2_sepconv2", new SeparableConvolution2D.Builder(3,3).nOut(128).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "block2_sepconv1_act")
.addLayer("block2_sepconv2_bn", new BatchNormalization(), "block2_sepconv2")
.addLayer("block2_pool", new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.MAX).kernelSize(3,3).stride(2,2)
.convolutionMode(ConvolutionMode.Same).build(), "block2_sepconv2_bn")
.addVertex("add1", new ElementWiseVertex(ElementWiseVertex.Op.Add), "block2_pool", "residual1")
// residual2
.addLayer("residual2_conv", new ConvolutionLayer.Builder(1,1).stride(2,2).nOut(256).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "add1")
.addLayer("residual2", new BatchNormalization(), "residual2_conv")
// block3
.addLayer("block3_sepconv1_act", new ActivationLayer(Activation.RELU), "add1")
.addLayer("block3_sepconv1", new SeparableConvolution2D.Builder(3,3).nOut(256).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "block3_sepconv1_act")
.addLayer("block3_sepconv1_bn", new BatchNormalization(), "block3_sepconv1")
.addLayer("block3_sepconv2_act", new ActivationLayer(Activation.RELU), "block3_sepconv1_bn")
.addLayer("block3_sepconv2", new SeparableConvolution2D.Builder(3,3).nOut(256).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "block3_sepconv2_act")
.addLayer("block3_sepconv2_bn", new BatchNormalization(), "block3_sepconv2")
.addLayer("block3_pool", new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.MAX).kernelSize(3,3).stride(2,2)
.convolutionMode(ConvolutionMode.Same).build(), "block3_sepconv2_bn")
.addVertex("add2", new ElementWiseVertex(ElementWiseVertex.Op.Add), "block3_pool", "residual2")
// residual3
.addLayer("residual3_conv", new ConvolutionLayer.Builder(1,1).stride(2,2).nOut(728).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "add1")
.addLayer("residual3", new BatchNormalization(), "residual3_conv")
// block4
.addLayer("block4_sepconv1_act", new ActivationLayer(Activation.RELU), "add2")
.addLayer("block4_sepconv1", new SeparableConvolution2D.Builder(3,3).nOut(728).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "block4_sepconv1_act")
.addLayer("block4_sepconv1_bn", new BatchNormalization(), "block4_sepconv1")
.addLayer("block4_sepconv2_act", new ActivationLayer(Activation.RELU), "block4_sepconv1_bn")
.addLayer("block4_sepconv2", new SeparableConvolution2D.Builder(3,3).nOut(728).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "block4_sepconv2_act")
.addLayer("block4_sepconv2_bn", new BatchNormalization(), "block4_sepconv2")
.addLayer("block4_pool", new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.MAX).kernelSize(3,3).stride(2,2)
.convolutionMode(ConvolutionMode.Same).build(), "block4_sepconv2_bn")
.addVertex("add3", new ElementWiseVertex(ElementWiseVertex.Op.Add), "block4_pool", "residual3");
// towers
int residual = 3;
int block = 5;
for(int i = 0; i < 8; i++) {
String previousInput = "add"+residual;
String blockName = "block"+block;
graph
.addLayer(blockName+"_sepconv1_act", new ActivationLayer(Activation.RELU), previousInput)
.addLayer(blockName+"_sepconv1", new SeparableConvolution2D.Builder(3,3).nOut(728).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), blockName+"_sepconv1_act")
.addLayer(blockName+"_sepconv1_bn", new BatchNormalization(), blockName+"_sepconv1")
.addLayer(blockName+"_sepconv2_act", new ActivationLayer(Activation.RELU), blockName+"_sepconv1_bn")
.addLayer(blockName+"_sepconv2", new SeparableConvolution2D.Builder(3,3).nOut(728).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), blockName+"_sepconv2_act")
.addLayer(blockName+"_sepconv2_bn", new BatchNormalization(), blockName+"_sepconv2")
.addLayer(blockName+"_sepconv3_act", new ActivationLayer(Activation.RELU), blockName+"_sepconv2_bn")
.addLayer(blockName+"_sepconv3", new SeparableConvolution2D.Builder(3,3).nOut(728).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), blockName+"_sepconv3_act")
.addLayer(blockName+"_sepconv3_bn", new BatchNormalization(), blockName+"_sepconv3")
.addVertex("add"+(residual+1), new ElementWiseVertex(ElementWiseVertex.Op.Add), blockName+"_sepconv3_bn", previousInput);
residual += 1;
block += 1;
}
graph
// residual11
.addLayer("residual11_conv", new ConvolutionLayer.Builder(1,1).stride(2,2).nOut(1024).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "add10")
.addLayer("residual11", new BatchNormalization(), "residual11_conv")
// block13
.addLayer("block13_sepconv1_act", new ActivationLayer(Activation.RELU), "add10")
.addLayer("block13_sepconv1", new SeparableConvolution2D.Builder(3,3).nOut(728).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "block13_sepconv1_act")
.addLayer("block13_sepconv1_bn", new BatchNormalization(), "block13_sepconv1")
.addLayer("block13_sepconv2_act", new ActivationLayer(Activation.RELU), "block13_sepconv1_bn")
.addLayer("block13_sepconv2", new SeparableConvolution2D.Builder(3,3).nOut(1024).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "block13_sepconv2_act")
.addLayer("block13_sepconv2_bn", new BatchNormalization(), "block13_sepconv2")
.addLayer("block13_pool", new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.MAX).kernelSize(3,3).stride(2,2)
.convolutionMode(ConvolutionMode.Same).build(), "block13_sepconv2_bn")
.addVertex("add11", new ElementWiseVertex(ElementWiseVertex.Op.Add), "block13_pool", "residual11")
// block14
.addLayer("block14_sepconv1", new SeparableConvolution2D.Builder(3,3).nOut(1536).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "add11")
.addLayer("block14_sepconv1_bn", new BatchNormalization(), "block14_sepconv1")
.addLayer("block14_sepconv1_act", new ActivationLayer(Activation.RELU), "block14_sepconv1_bn")
.addLayer("block14_sepconv2", new SeparableConvolution2D.Builder(3,3).nOut(2048).hasBias(false)
.convolutionMode(ConvolutionMode.Same).cudnnAlgoMode(cudnnAlgoMode).build(), "add11")
.addLayer("block14_sepconv2_bn", new BatchNormalization(), "block14_sepconv2")
.addLayer("block14_sepconv2_act", new ActivationLayer(Activation.RELU), "block14_sepconv2_bn")
.addLayer("avg_pool", new GlobalPoolingLayer.Builder(PoolingType.AVG).build(), "block14_sepconv2_act")
.addLayer("predictions", new OutputLayer.Builder(LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).build(), "avg_pool")
.setOutputs("predictions")
.backprop(true)
.pretrain(false);
return graph;
}
@Override
public ModelMetaData metaData() {
return new ModelMetaData(new int[][] {inputShape}, 1, ZooType.CNN);
}
@Override
public void setInputShape(int[][] inputShape) {
this.inputShape = inputShape[0];
}
}
© 2015 - 2024 Weber Informatics LLC | Privacy Policy