Last active
September 3, 2018 09:33
-
-
Save AlexDBlack/916dc4dd2a3df5d5b3241a0f60ff2285 to your computer and use it in GitHub Desktop.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
package org.deeplearning4j; | |
import org.deeplearning4j.nn.api.OptimizationAlgorithm; | |
import org.deeplearning4j.nn.conf.*; | |
import org.deeplearning4j.nn.conf.graph.rnn.DuplicateToTimeSeriesVertex; | |
import org.deeplearning4j.nn.conf.graph.rnn.LastTimeStepVertex; | |
import org.deeplearning4j.nn.conf.inputs.InputType; | |
import org.deeplearning4j.nn.conf.layers.ConvolutionLayer; | |
import org.deeplearning4j.nn.conf.layers.LSTM; | |
import org.deeplearning4j.nn.conf.layers.RnnOutputLayer; | |
import org.deeplearning4j.nn.conf.layers.SubsamplingLayer; | |
import org.deeplearning4j.nn.conf.preprocessor.CnnToRnnPreProcessor; | |
import org.deeplearning4j.nn.conf.preprocessor.RnnToCnnPreProcessor; | |
import org.deeplearning4j.nn.graph.ComputationGraph; | |
import org.deeplearning4j.nn.weights.WeightInit; | |
import org.junit.Test; | |
import org.nd4j.linalg.activations.Activation; | |
import org.nd4j.linalg.api.ndarray.INDArray; | |
import org.nd4j.linalg.factory.Nd4j; | |
import org.nd4j.linalg.learning.config.AdaGrad; | |
import org.nd4j.linalg.lossfunctions.LossFunctions; | |
import org.nd4j.linalg.schedule.MapSchedule; | |
import org.nd4j.linalg.schedule.ScheduleType; | |
import java.util.HashMap; | |
import java.util.Map; | |
public class Debug6316 { | |
@Test | |
public void test(){ | |
ComputationGraph cg = getComputationGraph(); | |
INDArray in = Nd4j.create(1, 501*501, 3); | |
INDArray label = Nd4j.create(1, 501*501, 3); | |
INDArray inMask = Nd4j.ones(1, 3); | |
INDArray lMask = Nd4j.ones(1, 3); | |
cg.fit(new INDArray[]{in}, new INDArray[]{label}, new INDArray[]{inMask}, new INDArray[]{lMask}); | |
} | |
public static ComputationGraph getComputationGraph(){ | |
ComputationGraph multiLayerNetwork; | |
NeuralNetConfiguration.Builder builder = new NeuralNetConfiguration.Builder(); | |
builder.seed(140); | |
builder.optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT); | |
builder.weightInit(WeightInit.XAVIER); | |
Map<Integer, Double> lrSchedule = new HashMap<>(); | |
lrSchedule.put(0, 1e-2); | |
lrSchedule.put(400, 1e-3); | |
lrSchedule.put(3000, 1e-4); | |
MapSchedule mapSchedule = new MapSchedule(ScheduleType.ITERATION, lrSchedule); | |
builder.updater(new AdaGrad(mapSchedule)); | |
int lstmHiddenCount = 200; | |
int cnnStride1 = 5; | |
int kernelSize1 = 3; | |
int cnnStride3 = 5; | |
int kernelSize3 = 3; | |
int channels = 1; | |
int padding = 1; | |
int samplingSize = 1; | |
int samplingStride = 1; | |
int cnn1Output = (501 - kernelSize1 + padding) / cnnStride1 + 1; | |
int cnn2Output = (cnn1Output - samplingSize + 0) / samplingStride + 1; | |
int cnn3Output = (cnn2Output - kernelSize3 + padding) / cnnStride3 + 1; | |
int lstmInWidth = cnn3Output; | |
int cnn4Output = (cnn3Output - samplingSize + 0) / samplingStride + 1; | |
int cnn5Output = (cnn4Output - kernelSize3 + padding) / cnnStride3 + 1; | |
lstmInWidth = cnn5Output + 1; // output of cnn | |
Map<String, InputPreProcessor> inputPreProcessors = new HashMap<String, InputPreProcessor>(); | |
inputPreProcessors.put("cnn1", new RnnToCnnPreProcessor(501, 501, channels)); | |
inputPreProcessors.put("lstm1", new CnnToRnnPreProcessor(lstmInWidth, lstmInWidth, 128)); | |
ComputationGraphConfiguration.GraphBuilder graphBuilder = builder.graphBuilder().pretrain(false).backprop(true) | |
.backpropType(BackpropType.Standard) | |
.addInputs("inputs") | |
// cnn | |
.addLayer("cnn1", | |
new ConvolutionLayer.Builder(new int[] { kernelSize1, kernelSize1 }, | |
new int[] { cnnStride1, cnnStride1 }, | |
new int[] { padding, padding }) | |
.nIn(channels) | |
.nOut(501) | |
.gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue) | |
.gradientNormalizationThreshold(10) | |
.updater(new AdaGrad(mapSchedule)) | |
.weightInit(WeightInit.RELU) | |
.activation(Activation.RELU).build(), "inputs") | |
// Output: (501 - kernelSize + padding) / cnn1Stride + 1 = 125 --> x * x * nOut = paramsNum | |
.addLayer("cnn2", new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.MAX, | |
new int[] { samplingSize, samplingSize }, new int[] { samplingStride, samplingStride }).build(), "cnn1") | |
// Output: (125-1+0)/1+1 = 125 | |
.addLayer("cnn3", | |
new ConvolutionLayer.Builder(new int[] { kernelSize3, kernelSize3 }, | |
new int[] { cnnStride3, cnnStride3 }, | |
new int[] { padding, padding }) | |
.nIn(501) | |
.nOut(128) | |
.gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue) | |
.gradientNormalizationThreshold(10) | |
.updater(new AdaGrad(mapSchedule)) | |
.weightInit(WeightInit.RELU) | |
.activation(Activation.RELU).build(), "cnn2"); | |
// Output: (125 - kernelSize + padding) / cnn3Stride + 1 = 25 --> x * x * 100 = ? | |
graphBuilder = graphBuilder.addLayer("cnn4", new SubsamplingLayer.Builder( | |
SubsamplingLayer.PoolingType.MAX, | |
new int[] { samplingSize, samplingSize }, | |
new int[] { samplingStride, samplingStride }).build(), "cnn3") | |
.addLayer("cnn5", | |
new ConvolutionLayer.Builder(new int[] { kernelSize3, kernelSize3 }, | |
new int[] { cnnStride3, cnnStride3 }, | |
new int[] { padding, padding }) | |
.nIn(128) | |
.nOut(128) | |
.updater(new AdaGrad(mapSchedule)) | |
.weightInit(WeightInit.RELU) | |
.activation(Activation.RELU).build(), "cnn4"); | |
graphBuilder = graphBuilder.addLayer("lstm1", new LSTM.Builder() | |
.activation(Activation.SOFTSIGN) | |
.nIn(lstmInWidth * lstmInWidth * 128) | |
.nOut(lstmHiddenCount) | |
.gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue) | |
.gradientNormalizationThreshold(10) | |
.updater(new AdaGrad(mapSchedule)) | |
.build(), "cnn5"); | |
graphBuilder = graphBuilder.addVertex("thoughtVector", new LastTimeStepVertex("inputs"), "lstm1"); | |
graphBuilder = graphBuilder.addVertex("dup", new DuplicateToTimeSeriesVertex("inputs"), "thoughtVector"); | |
graphBuilder = graphBuilder.addLayer("lstmDecode1", new LSTM.Builder() | |
.activation(Activation.SOFTSIGN) | |
.nIn(lstmHiddenCount) | |
.nOut(lstmHiddenCount) | |
.gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue) | |
.gradientNormalizationThreshold(10) | |
.updater(new AdaGrad(mapSchedule)) | |
.build(), "dup") | |
.addLayer("output", new RnnOutputLayer | |
.Builder(LossFunctions.LossFunction.MSE) | |
.activation(Activation.RELU) | |
.nIn(lstmHiddenCount) | |
.nOut(501 * 501) | |
.gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue) | |
.gradientNormalizationThreshold(10) | |
.updater(new AdaGrad(mapSchedule)) | |
.build(), "lstmDecode1"); | |
graphBuilder = graphBuilder.setOutputs("output"); | |
graphBuilder.setInputPreProcessors(inputPreProcessors); | |
int inputSize = 30 * 2; | |
graphBuilder.setInputTypes(InputType.recurrent(501 * 501, inputSize)); | |
multiLayerNetwork = new ComputationGraph(graphBuilder.build()); | |
multiLayerNetwork.init(); | |
return multiLayerNetwork; | |
} | |
} |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment