org.deeplearning4j.nn.transferlearning.TransferLearning Java Examples
The following examples show how to use
org.deeplearning4j.nn.transferlearning.TransferLearning.
You can vote up the ones you like or vote down the ones you don't like,
and go to the original project or source file by following the links above each example. You may check out the related API usage on the sidebar.
Example #1
| Source File: TestInstantiation.java From deeplearning4j with Apache License 2.0 | 6 votes |
|
@Test
public void testInitPretrained() throws IOException {
ignoreIfCuda();
ZooModel model = ResNet50.builder().numClasses(0).build(); //num labels doesn't matter since we're getting pretrained imagenet
assertTrue(model.pretrainedAvailable(PretrainedType.IMAGENET));
ComputationGraph initializedModel = (ComputationGraph) model.initPretrained();
INDArray f = Nd4j.rand(new int[]{1, 3, 224, 224});
INDArray[] result = initializedModel.output(f);
assertArrayEquals(result[0].shape(), new long[]{1, 1000});
//Test fitting. Not ewe need to use transfer learning, as ResNet50 has a dense layer, not an OutputLayer
initializedModel = new TransferLearning.GraphBuilder(initializedModel)
.removeVertexAndConnections("fc1000")
.addLayer("fc1000", new OutputLayer.Builder()
.lossFunction(LossFunctions.LossFunction.MCXENT)
.nIn(2048).nOut(1000).activation(Activation.SOFTMAX).build(), "flatten_1")
.setOutputs("fc1000")
.build();
initializedModel.fit(new org.nd4j.linalg.dataset.DataSet(f, TestUtils.randomOneHot(1, 1000, 12345)));
}
Example #2
| Source File: TestTransferStatsCollection.java From deeplearning4j with Apache License 2.0 | 6 votes |
|
@Test
public void test() throws IOException {
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().list()
.layer(0, new DenseLayer.Builder().nIn(10).nOut(10).build())
.layer(1, new OutputLayer.Builder().activation(Activation.SOFTMAX).nIn(10).nOut(10).build()).build();
MultiLayerNetwork net = new MultiLayerNetwork(conf);
net.init();
MultiLayerNetwork net2 =
new TransferLearning.Builder(net)
.fineTuneConfiguration(
new FineTuneConfiguration.Builder().updater(new Sgd(0.01)).build())
.setFeatureExtractor(0).build();
net2.setListeners(new StatsListener(new InMemoryStatsStorage()));
//Previosuly: failed on frozen layers
net2.fit(new DataSet(Nd4j.rand(8, 10), Nd4j.rand(8, 10)));
}
Example #3
| Source File: VGG16.java From Java-Machine-Learning-for-Computer-Vision with MIT License | 6 votes |
|
public static ComputationGraph loadModel() throws IOException {
ZooModel zooModel = new org.deeplearning4j.zoo.model.VGG16();
ComputationGraph vgg16 = (ComputationGraph) zooModel.initPretrained(PretrainedType.IMAGENET);
ComputationGraph vgg16Transfer = new TransferLearning.GraphBuilder(vgg16)
.removeVertexKeepConnections("block5_pool")
.addLayer("block5_pool", avgPoolLayer(), "block5_conv3")
.removeVertexKeepConnections("block4_pool")
.addLayer("block4_pool", avgPoolLayer(), "block4_conv3")
.removeVertexKeepConnections("block3_pool")
.addLayer("block3_pool", avgPoolLayer(), "block3_conv3")
.removeVertexKeepConnections("block2_pool")
.addLayer("block2_pool", avgPoolLayer(), "block2_conv2")
.removeVertexKeepConnections("block1_pool")
.addLayer("block1_pool", avgPoolLayer(), "block1_conv2")
.setInputTypes(InputType
.convolutionalFlat(NeuralStyleTransfer.HEIGHT, NeuralStyleTransfer.WIDTH, NeuralStyleTransfer.CHANNELS))
.removeVertexAndConnections("fc2")
.removeVertexAndConnections("fc1")
.removeVertexAndConnections("flatten")
.removeVertexAndConnections("predictions")
.setOutputs("block5_pool")
.build();
vgg16Transfer.initGradientsView();
System.out.println("vgg16Transfer.summary() = " + vgg16Transfer.summary());
return vgg16Transfer;
}
Example #4
| Source File: MiscTests.java From deeplearning4j with Apache License 2.0 | 5 votes |
|
@Test
public void testTransferVGG() throws Exception {
//https://github.com/deeplearning4j/deeplearning4j/issues/5167
DataSet ds = new DataSet();
ds.setFeatures(Nd4j.create(1, 3, 224, 224));
ds.setLabels(Nd4j.create(1, 2));
ComputationGraph model = (ComputationGraph)(
VGG16.builder().build()
.initPretrained(PretrainedType.IMAGENET));
// System.out.println(model.summary());
ComputationGraph transferModel = new TransferLearning.GraphBuilder(model)
.setFeatureExtractor("fc2")
.removeVertexKeepConnections("predictions")
.addLayer("predictions",
new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
.nIn(4096).nOut(2)
.weightInit(WeightInit.XAVIER)
.activation(Activation.SOFTMAX).build(), "fc2")
.build();
// System.out.println(transferModel.summary());
// System.out.println("Fitting");
transferModel.fit(ds);
ComputationGraph g2 = TestUtils.testModelSerialization(transferModel);
g2.fit(ds);
}
Example #5
| Source File: TestInstantiation.java From deeplearning4j with Apache License 2.0 | 5 votes |
|
@Test
public void testTransferLearning() throws Exception {
ignoreIfCuda();
//https://github.com/deeplearning4j/deeplearning4j/issues/7193
ComputationGraph cg = (ComputationGraph) ResNet50.builder().build().initPretrained();
cg = new TransferLearning.GraphBuilder(cg)
.addLayer("out", new LossLayer.Builder().lossFunction(LossFunctions.LossFunction.MCXENT).activation(Activation.IDENTITY).build(), "fc1000")
.setInputTypes(InputType.convolutional(224, 224, 3))
.setOutputs("out")
.build();
}
Example #6
| Source File: AbstractZooModel.java From wekaDeeplearning4j with GNU General Public License v3.0 | 5 votes |
|
/**
* Checks if we need to add a final output layer - also applies pooling beforehand if necessary
* @param computationGraph Input ComputationGraph
* @return Finalized ComputationGraph
*/
protected ComputationGraph addFinalOutputLayer(ComputationGraph computationGraph) {
org.deeplearning4j.nn.conf.layers.Layer lastLayer = computationGraph.getLayers()[computationGraph.getNumLayers() - 1].conf().getLayer();
if (!Dl4jMlpClassifier.noOutputLayer(filterMode, lastLayer)) {
log.debug("No need to add output layer, ignoring");
return computationGraph;
}
try {
TransferLearning.GraphBuilder graphBuilder;
if (requiresPooling)
graphBuilder = new TransferLearning.GraphBuilder(computationGraph)
.fineTuneConfiguration(getFineTuneConfig())
.addLayer("intermediate_pooling", new GlobalPoolingLayer.Builder().build(), m_featureExtractionLayer)
.addLayer(m_predictionLayerName, createOutputLayer(), "intermediate_pooling")
.setOutputs(m_predictionLayerName);
else
graphBuilder = new TransferLearning.GraphBuilder(computationGraph)
.fineTuneConfiguration(getFineTuneConfig())
.addLayer(m_predictionLayerName, createOutputLayer(), m_featureExtractionLayer)
.setOutputs(m_predictionLayerName);
// Remove the old output layer, but keep the connections
graphBuilder.removeVertexKeepConnections(m_outputLayer);
// Remove any other layers we don't want
for (String layer : m_extraLayersToRemove) {
graphBuilder.removeVertexAndConnections(layer);
}
log.debug("Finished adding output layer");
return graphBuilder.build();
} catch (Exception ex) {
ex.printStackTrace();
log.error(computationGraph.summary());
return computationGraph;
}
}
Example #7
| Source File: AbstractZooModel.java From wekaDeeplearning4j with GNU General Public License v3.0 | 5 votes |
|
/**
* Checks if we need to add a final output layer - also applies pooling beforehand if necessary
* @param computationGraph Input ComputationGraph
* @return Finalized ComputationGraph
*/
protected ComputationGraph addFinalOutputLayer(ComputationGraph computationGraph) {
org.deeplearning4j.nn.conf.layers.Layer lastLayer = computationGraph.getLayers()[computationGraph.getNumLayers() - 1].conf().getLayer();
if (!Dl4jMlpClassifier.noOutputLayer(filterMode, lastLayer)) {
log.debug("No need to add output layer, ignoring");
return computationGraph;
}
try {
TransferLearning.GraphBuilder graphBuilder;
if (requiresPooling)
graphBuilder = new TransferLearning.GraphBuilder(computationGraph)
.fineTuneConfiguration(getFineTuneConfig())
.addLayer("intermediate_pooling", new GlobalPoolingLayer.Builder().build(), m_featureExtractionLayer)
.addLayer(m_predictionLayerName, createOutputLayer(), "intermediate_pooling")
.setOutputs(m_predictionLayerName);
else
graphBuilder = new TransferLearning.GraphBuilder(computationGraph)
.fineTuneConfiguration(getFineTuneConfig())
.addLayer(m_predictionLayerName, createOutputLayer(), m_featureExtractionLayer)
.setOutputs(m_predictionLayerName);
// Remove the old output layer, but keep the connections
graphBuilder.removeVertexKeepConnections(m_outputLayer);
// Remove any other layers we don't want
for (String layer : m_extraLayersToRemove) {
graphBuilder.removeVertexAndConnections(layer);
}
log.debug("Finished adding output layer");
return graphBuilder.build();
} catch (Exception ex) {
ex.printStackTrace();
log.error(computationGraph.summary());
return computationGraph;
}
}
Example #8
| Source File: BatchNormalizationTest.java From deeplearning4j with Apache License 2.0 | 5 votes |
|
@Test
public void testBatchNorm() throws Exception {
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
.seed(12345)
.updater(new Adam(1e-3))
.activation(Activation.TANH)
.list()
.layer(new ConvolutionLayer.Builder().nOut(5).kernelSize(2, 2).build())
.layer(new BatchNormalization())
.layer(new ConvolutionLayer.Builder().nOut(5).kernelSize(2, 2).build())
.layer(new OutputLayer.Builder().activation(Activation.SOFTMAX).lossFunction(LossFunctions.LossFunction.MCXENT).nOut(10).build())
.setInputType(InputType.convolutionalFlat(28, 28, 1))
.build();
MultiLayerNetwork net = new MultiLayerNetwork(conf);
net.init();
DataSetIterator iter = new EarlyTerminationDataSetIterator(new MnistDataSetIterator(32, true, 12345), 10);
net.fit(iter);
MultiLayerNetwork net2 = new TransferLearning.Builder(net)
.fineTuneConfiguration(FineTuneConfiguration.builder()
.updater(new AdaDelta())
.build())
.removeOutputLayer()
.addLayer(new BatchNormalization.Builder().nOut(3380).build())
.addLayer(new OutputLayer.Builder().activation(Activation.SOFTMAX).lossFunction(LossFunctions.LossFunction.MCXENT).nIn(3380).nOut(10).build())
.build();
net2.fit(iter);
}
Example #9
| Source File: TestComputationGraphNetwork.java From deeplearning4j with Apache License 2.0 | 5 votes |
|
@Test
public void printSummary() {
NeuralNetConfiguration.Builder overallConf = new NeuralNetConfiguration.Builder().updater(new Sgd(0.1))
.activation(Activation.IDENTITY);
ComputationGraphConfiguration conf = overallConf.graphBuilder().addInputs("inCentre", "inRight")
.addLayer("denseCentre0", new DenseLayer.Builder().nIn(10).nOut(9).build(), "inCentre")
.addLayer("denseCentre1", new DenseLayer.Builder().nIn(9).nOut(8).build(), "denseCentre0")
.addLayer("denseCentre2", new DenseLayer.Builder().nIn(8).nOut(7).build(), "denseCentre1")
.addLayer("denseCentre3", new DenseLayer.Builder().nIn(7).nOut(7).build(), "denseCentre2")
.addLayer("outCentre",
new OutputLayer.Builder(LossFunctions.LossFunction.MSE).nIn(7).nOut(4).build(),
"denseCentre3")
.addVertex("subsetLeft", new SubsetVertex(0, 3), "denseCentre1")
.addLayer("denseLeft0", new DenseLayer.Builder().nIn(4).nOut(5).build(), "subsetLeft")
.addLayer("outLeft",
new OutputLayer.Builder(LossFunctions.LossFunction.MSE).nIn(5).nOut(6).build(),
"denseLeft0")
.addLayer("denseRight", new DenseLayer.Builder().nIn(7).nOut(7).build(), "denseCentre2")
.addLayer("denseRight0", new DenseLayer.Builder().nIn(2).nOut(3).build(), "inRight")
.addVertex("mergeRight", new MergeVertex(), "denseRight", "denseRight0")
.addLayer("denseRight1", new DenseLayer.Builder().nIn(10).nOut(5).build(), "mergeRight")
.addLayer("outRight",
new OutputLayer.Builder(LossFunctions.LossFunction.MSE).nIn(5).nOut(5).build(),
"denseRight1")
.setOutputs("outLeft", "outCentre", "outRight").build();
ComputationGraph modelToTune = new ComputationGraph(conf);
modelToTune.init();
// System.out.println(modelToTune.summary());
modelToTune.summary();
ComputationGraph modelNow =
new TransferLearning.GraphBuilder(modelToTune).setFeatureExtractor("denseCentre2").build();
// System.out.println(modelNow.summary());
// System.out.println(modelNow.summary(InputType.feedForward(10),InputType.feedForward(2)));
modelNow.summary();
modelNow.summary(InputType.feedForward(10),InputType.feedForward(2));
}
Example #10
| Source File: FrozenLayerTest.java From deeplearning4j with Apache License 2.0 | 4 votes |
|
@Test
public void testFrozenCompGraph() {
DataSet randomData = new DataSet(Nd4j.rand(10, 4), Nd4j.rand(10, 3));
NeuralNetConfiguration.Builder overallConf = new NeuralNetConfiguration.Builder().updater(new Sgd(0.1))
.activation(Activation.IDENTITY);
ComputationGraph modelToFineTune = new ComputationGraph(overallConf.graphBuilder().addInputs("layer0In")
.addLayer("layer0", new DenseLayer.Builder().nIn(4).nOut(3).build(), "layer0In")
.addLayer("layer1", new DenseLayer.Builder().nIn(3).nOut(2).build(), "layer0")
.addLayer("layer2", new DenseLayer.Builder().nIn(2).nOut(3).build(), "layer1")
.addLayer("layer3",
new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(
LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nIn(3).nOut(3)
.build(),
"layer2")
.setOutputs("layer3").build());
modelToFineTune.init();
INDArray asFrozenFeatures = modelToFineTune.feedForward(randomData.getFeatures(), false).get("layer1");
ComputationGraph modelNow =
new TransferLearning.GraphBuilder(modelToFineTune).setFeatureExtractor("layer1").build();
ComputationGraph notFrozen = new ComputationGraph(overallConf.graphBuilder().addInputs("layer0In")
.addLayer("layer0", new DenseLayer.Builder().nIn(2).nOut(3).build(), "layer0In")
.addLayer("layer1",
new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(
LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nIn(3).nOut(3)
.build(),
"layer0")
.setOutputs("layer1").build());
notFrozen.init();
notFrozen.setParams(Nd4j.hstack(modelToFineTune.getLayer("layer2").params(),
modelToFineTune.getLayer("layer3").params()));
int i = 0;
while (i < 5) {
notFrozen.fit(new DataSet(asFrozenFeatures, randomData.getLabels()));
modelNow.fit(randomData);
i++;
}
assertEquals(Nd4j.hstack(modelToFineTune.getLayer("layer0").params(),
modelToFineTune.getLayer("layer1").params(), notFrozen.params()), modelNow.params());
}
Example #11
| Source File: TestComputationGraphNetwork.java From deeplearning4j with Apache License 2.0 | 4 votes |
|
@Test
public void testSummary() {
int V_WIDTH = 130;
int V_HEIGHT = 130;
int V_NFRAMES = 150;
ComputationGraphConfiguration confForArchitecture =
new NeuralNetConfiguration.Builder().seed(12345).l2(0.001) //l2 regularization on all layers
.updater(new AdaGrad(0.4)).graphBuilder()
.addInputs("in")
.addLayer("layer0", new ConvolutionLayer.Builder(10, 10).nIn(3) //3 channels: RGB
.nOut(30).stride(4, 4).activation(Activation.RELU).weightInit(
WeightInit.RELU).build(),"in") //Output: (130-10+0)/4+1 = 31 -> 31*31*30
.addLayer("layer1", new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.MAX)
.kernelSize(3, 3).stride(2, 2).build(),"layer0") //(31-3+0)/2+1 = 15
.addLayer("layer2", new ConvolutionLayer.Builder(3, 3).nIn(30).nOut(10).stride(2, 2)
.activation(Activation.RELU).weightInit(WeightInit.RELU)
.updater(Updater.ADAGRAD).build(), "layer1") //Output: (15-3+0)/2+1 = 7 -> 7*7*10 = 490
.addLayer("layer3", new DenseLayer.Builder().activation(Activation.RELU).nIn(490).nOut(50)
.weightInit(WeightInit.RELU).gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue)
.gradientNormalizationThreshold(10).build(), "layer2")
.addLayer("layer4", new GravesLSTM.Builder().activation(Activation.SOFTSIGN).nIn(50)
.nOut(50).weightInit(WeightInit.XAVIER).updater(Updater.ADAGRAD)
.gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue)
.gradientNormalizationThreshold(10)
.build(), "layer3")
.addLayer("layer5", new RnnOutputLayer.Builder(LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nIn(50).nOut(4) //4 possible shapes: circle, square, arc, line
.weightInit(WeightInit.XAVIER)
.gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue)
.gradientNormalizationThreshold(10).build(), "layer4")
.setOutputs("layer5")
.inputPreProcessor("layer0", new RnnToCnnPreProcessor(V_HEIGHT, V_WIDTH, 3))
.inputPreProcessor("layer3", new CnnToFeedForwardPreProcessor(7, 7, 10))
.inputPreProcessor("layer4", new FeedForwardToRnnPreProcessor())
.backpropType(BackpropType.TruncatedBPTT)
.tBPTTForwardLength(V_NFRAMES / 5).tBPTTBackwardLength(V_NFRAMES / 5).build();
ComputationGraph modelExpectedArch = new ComputationGraph(confForArchitecture);
modelExpectedArch.init();
ComputationGraph modelMow = new TransferLearning.GraphBuilder(modelExpectedArch).setFeatureExtractor("layer2").build();
// System.out.println(modelExpectedArch.summary());
// System.out.println(modelMow.summary());
// System.out.println(modelExpectedArch.summary(InputType.recurrent(V_HEIGHT* V_WIDTH* 3)));
modelExpectedArch.summary();
modelMow.summary();
modelExpectedArch.summary(InputType.recurrent(V_HEIGHT* V_WIDTH* 3));
}
Example #12
| Source File: TestGraphNodes.java From deeplearning4j with Apache License 2.0 | 4 votes |
|
@Test
public void testLastTimeStepWithTransfer(){
int lstmLayerSize = 16;
int numLabelClasses = 10;
int numInputs = 5;
ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder()
.trainingWorkspaceMode(WorkspaceMode.NONE)
.inferenceWorkspaceMode(WorkspaceMode.NONE)
.seed(123) //Random number generator seed for improved repeatability. Optional.
.updater(new AdaDelta())
.weightInit(WeightInit.XAVIER)
.graphBuilder()
.addInputs("rr")
.setInputTypes(InputType.recurrent(30))
.addLayer("1", new GravesLSTM.Builder().activation(Activation.TANH).nIn(numInputs).nOut(lstmLayerSize).dropOut(0.9).build(), "rr")
.addLayer("2", new RnnOutputLayer.Builder(LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nOut(numLabelClasses).build(), "1")
.setOutputs("2")
.build();
ComputationGraph net = new ComputationGraph(conf);
net.init();
ComputationGraph updatedModel = new TransferLearning.GraphBuilder(net)
.addVertex("laststepoutput", new LastTimeStepVertex("rr"), "2")
.setOutputs("laststepoutput")
.build();
INDArray input = Nd4j.rand(new int[]{10, numInputs, 16});
INDArray[] out = updatedModel.output(input);
assertNotNull(out);
assertEquals(1, out.length);
assertNotNull(out[0]);
assertArrayEquals(new long[]{10, numLabelClasses}, out[0].shape());
Map<String,INDArray> acts = updatedModel.feedForward(input, false);
assertEquals(4, acts.size()); //2 layers + input + vertex output
assertNotNull(acts.get("laststepoutput"));
assertArrayEquals(new long[]{10, numLabelClasses}, acts.get("laststepoutput").shape());
String toString = out[0].toString();
}
Example #13
| Source File: MultiLayerTest.java From deeplearning4j with Apache License 2.0 | 4 votes |
|
@Test
public void testSummary() {
int V_WIDTH = 130;
int V_HEIGHT = 130;
int V_NFRAMES = 150;
MultiLayerConfiguration confForArchitecture =
new NeuralNetConfiguration.Builder().seed(12345).l2(0.001) //l2 regularization on all layers
.optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
.list()
.layer(0, new ConvolutionLayer.Builder(10, 10).nIn(3) //3 channels: RGB
.nOut(30).stride(4, 4).activation(Activation.RELU).weightInit(
WeightInit.RELU)
.updater(Updater.ADAGRAD).build()) //Output: (130-10+0)/4+1 = 31 -> 31*31*30
.layer(1, new SubsamplingLayer.Builder(SubsamplingLayer.PoolingType.MAX)
.kernelSize(3, 3).stride(2, 2).build()) //(31-3+0)/2+1 = 15
.layer(2, new ConvolutionLayer.Builder(3, 3).nIn(30).nOut(10).stride(2, 2)
.activation(Activation.RELU).weightInit(WeightInit.RELU)
.updater(Updater.ADAGRAD).build()) //Output: (15-3+0)/2+1 = 7 -> 7*7*10 = 490
.layer(3, new DenseLayer.Builder().activation(Activation.RELU).nIn(490).nOut(50)
.weightInit(WeightInit.RELU).updater(Updater.ADAGRAD)
.gradientNormalization(
GradientNormalization.ClipElementWiseAbsoluteValue)
.gradientNormalizationThreshold(10).build())
.layer(4, new GravesLSTM.Builder().activation(Activation.SOFTSIGN).nIn(50)
.nOut(50).weightInit(WeightInit.XAVIER).updater(Updater.ADAGRAD)
.gradientNormalization(
GradientNormalization.ClipElementWiseAbsoluteValue)
.gradientNormalizationThreshold(10)
.build())
.layer(5, new RnnOutputLayer.Builder(LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nIn(50).nOut(4) //4 possible shapes: circle, square, arc, line
.updater(Updater.ADAGRAD).weightInit(WeightInit.XAVIER)
.gradientNormalization(
GradientNormalization.ClipElementWiseAbsoluteValue)
.gradientNormalizationThreshold(10).build())
.inputPreProcessor(0, new RnnToCnnPreProcessor(V_HEIGHT, V_WIDTH, 3))
.inputPreProcessor(3, new CnnToFeedForwardPreProcessor(7, 7, 10))
.inputPreProcessor(4, new FeedForwardToRnnPreProcessor())
.backpropType(BackpropType.TruncatedBPTT)
.tBPTTForwardLength(V_NFRAMES / 5).tBPTTBackwardLength(V_NFRAMES / 5).build();
MultiLayerNetwork modelExpectedArch = new MultiLayerNetwork(confForArchitecture);
modelExpectedArch.init();
MultiLayerNetwork modelMow = new TransferLearning.Builder(modelExpectedArch).setFeatureExtractor(2).build();
// System.out.println(modelExpectedArch.summary());
// System.out.println(modelMow.summary());
// System.out.println(modelMow.summary(InputType.recurrent(V_HEIGHT*V_WIDTH*3)));
}
Example #14
| Source File: FrozenLayerTest.java From deeplearning4j with Apache License 2.0 | 4 votes |
|
@Test
public void cloneCompGraphFrozen() {
DataSet randomData = new DataSet(Nd4j.rand(10, 4), Nd4j.rand(10, 3));
NeuralNetConfiguration.Builder overallConf = new NeuralNetConfiguration.Builder().updater(new Sgd(0.1))
.activation(Activation.IDENTITY);
ComputationGraph modelToFineTune = new ComputationGraph(overallConf.graphBuilder().addInputs("layer0In")
.addLayer("layer0", new DenseLayer.Builder().nIn(4).nOut(3).build(), "layer0In")
.addLayer("layer1", new DenseLayer.Builder().nIn(3).nOut(2).build(), "layer0")
.addLayer("layer2", new DenseLayer.Builder().nIn(2).nOut(3).build(), "layer1")
.addLayer("layer3",
new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(
LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nIn(3).nOut(3)
.build(),
"layer2")
.setOutputs("layer3").build());
modelToFineTune.init();
INDArray asFrozenFeatures = modelToFineTune.feedForward(randomData.getFeatures(), false).get("layer1");
ComputationGraph modelNow =
new TransferLearning.GraphBuilder(modelToFineTune).setFeatureExtractor("layer1").build();
ComputationGraph clonedModel = modelNow.clone();
//Check json
assertEquals(clonedModel.getConfiguration().toJson(), modelNow.getConfiguration().toJson());
//Check params
assertEquals(modelNow.params(), clonedModel.params());
ComputationGraph notFrozen = new ComputationGraph(overallConf.graphBuilder().addInputs("layer0In")
.addLayer("layer0", new DenseLayer.Builder().nIn(2).nOut(3).build(), "layer0In")
.addLayer("layer1",
new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(
LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nIn(3).nOut(3)
.build(),
"layer0")
.setOutputs("layer1").build());
notFrozen.init();
notFrozen.setParams(Nd4j.hstack(modelToFineTune.getLayer("layer2").params(),
modelToFineTune.getLayer("layer3").params()));
int i = 0;
while (i < 5) {
notFrozen.fit(new DataSet(asFrozenFeatures, randomData.getLabels()));
modelNow.fit(randomData);
clonedModel.fit(randomData);
i++;
}
INDArray expectedParams = Nd4j.hstack(modelToFineTune.getLayer("layer0").params(),
modelToFineTune.getLayer("layer1").params(), notFrozen.params());
assertEquals(expectedParams, modelNow.params());
assertEquals(expectedParams, clonedModel.params());
}
Example #15
| Source File: FrozenLayerTest.java From deeplearning4j with Apache License 2.0 | 4 votes |
|
@Test
public void cloneMLNFrozen() {
DataSet randomData = new DataSet(Nd4j.rand(10, 4), Nd4j.rand(10, 3));
NeuralNetConfiguration.Builder overallConf = new NeuralNetConfiguration.Builder().updater(new Sgd(0.1))
.activation(Activation.IDENTITY);
MultiLayerNetwork modelToFineTune = new MultiLayerNetwork(overallConf.list()
.layer(0, new DenseLayer.Builder().nIn(4).nOut(3).build())
.layer(1, new DenseLayer.Builder().nIn(3).nOut(2).build())
.layer(2, new DenseLayer.Builder().nIn(2).nOut(3).build())
.layer(3, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(
LossFunctions.LossFunction.MCXENT).activation(Activation.SOFTMAX).nIn(3).nOut(3)
.build())
.build());
modelToFineTune.init();
INDArray asFrozenFeatures = modelToFineTune.feedForwardToLayer(2, randomData.getFeatures(), false).get(2);
MultiLayerNetwork modelNow = new TransferLearning.Builder(modelToFineTune).setFeatureExtractor(1).build();
MultiLayerNetwork clonedModel = modelNow.clone();
//Check json
assertEquals(modelNow.getLayerWiseConfigurations().toJson(), clonedModel.getLayerWiseConfigurations().toJson());
//Check params
assertEquals(modelNow.params(), clonedModel.params());
MultiLayerNetwork notFrozen = new MultiLayerNetwork(
overallConf.list().layer(0, new DenseLayer.Builder().nIn(2).nOut(3).build())
.layer(1, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(
LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nIn(3).nOut(3)
.build())
.build(),
Nd4j.hstack(modelToFineTune.getLayer(2).params(), modelToFineTune.getLayer(3).params()));
int i = 0;
while (i < 5) {
notFrozen.fit(new DataSet(asFrozenFeatures, randomData.getLabels()));
modelNow.fit(randomData);
clonedModel.fit(randomData);
i++;
}
INDArray expectedParams = Nd4j.hstack(modelToFineTune.getLayer(0).params(),
modelToFineTune.getLayer(1).params(), notFrozen.params());
assertEquals(expectedParams, modelNow.params());
assertEquals(expectedParams, clonedModel.params());
}
Example #16
| Source File: FrozenLayerTest.java From deeplearning4j with Apache License 2.0 | 4 votes |
|
@Test
public void testFrozen() {
DataSet randomData = new DataSet(Nd4j.rand(10, 4), Nd4j.rand(10, 3));
NeuralNetConfiguration.Builder overallConf = new NeuralNetConfiguration.Builder().updater(new Sgd(0.1))
.activation(Activation.IDENTITY);
FineTuneConfiguration finetune = new FineTuneConfiguration.Builder().updater(new Sgd(0.1)).build();
MultiLayerNetwork modelToFineTune = new MultiLayerNetwork(overallConf.clone().list()
.layer(0, new DenseLayer.Builder().nIn(4).nOut(3).build())
.layer(1, new DenseLayer.Builder().nIn(3).nOut(2).build())
.layer(2, new DenseLayer.Builder().nIn(2).nOut(3).build())
.layer(3, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(
LossFunctions.LossFunction.MCXENT).activation(Activation.SOFTMAX).nIn(3).nOut(3)
.build())
.build());
modelToFineTune.init();
List<INDArray> ff = modelToFineTune.feedForwardToLayer(2, randomData.getFeatures(), false);
INDArray asFrozenFeatures = ff.get(2);
MultiLayerNetwork modelNow = new TransferLearning.Builder(modelToFineTune).fineTuneConfiguration(finetune)
.setFeatureExtractor(1).build();
INDArray paramsLastTwoLayers =
Nd4j.hstack(modelToFineTune.getLayer(2).params(), modelToFineTune.getLayer(3).params());
MultiLayerNetwork notFrozen = new MultiLayerNetwork(overallConf.clone().list()
.layer(0, new DenseLayer.Builder().nIn(2).nOut(3).build())
.layer(1, new org.deeplearning4j.nn.conf.layers.OutputLayer.Builder(
LossFunctions.LossFunction.MCXENT).activation(Activation.SOFTMAX).nIn(3).nOut(3)
.build())
.build(), paramsLastTwoLayers);
// assertEquals(modelNow.getLayer(2).conf(), notFrozen.getLayer(0).conf()); //Equal, other than names
// assertEquals(modelNow.getLayer(3).conf(), notFrozen.getLayer(1).conf()); //Equal, other than names
//Check: forward pass
INDArray outNow = modelNow.output(randomData.getFeatures());
INDArray outNotFrozen = notFrozen.output(asFrozenFeatures);
assertEquals(outNow, outNotFrozen);
for (int i = 0; i < 5; i++) {
notFrozen.fit(new DataSet(asFrozenFeatures, randomData.getLabels()));
modelNow.fit(randomData);
}
INDArray expected = Nd4j.hstack(modelToFineTune.getLayer(0).params(), modelToFineTune.getLayer(1).params(),
notFrozen.params());
INDArray act = modelNow.params();
assertEquals(expected, act);
}
Example #17
| Source File: TransferLearningVGG16.java From Java-Machine-Learning-for-Computer-Vision with MIT License | 4 votes |
|
public void train() throws IOException {
ComputationGraph preTrainedNet = loadVGG16PreTrainedWeights();
log.info("VGG 16 Architecture");
log.info(preTrainedNet.summary());
log.info("Start Downloading NeuralNetworkTrainingData...");
downloadAndUnzipDataForTheFirstTime();
log.info("NeuralNetworkTrainingData Downloaded and unzipped");
neuralNetworkTrainingData = new DataStorage() {
}.loadData();
FineTuneConfiguration fineTuneConf = new FineTuneConfiguration.Builder()
.learningRate(LEARNING_RATE)
.optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
.updater(Updater.NESTEROVS)
.seed(1234)
.build();
ComputationGraph vgg16Transfer = new TransferLearning.GraphBuilder(preTrainedNet)
.fineTuneConfiguration(fineTuneConf)
.setFeatureExtractor(FREEZE_UNTIL_LAYER)
.removeVertexKeepConnections("predictions")
.addLayer("predictions",
new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
.nIn(4096)
.nOut(NUM_POSSIBLE_LABELS)
.weightInit(WeightInit.XAVIER)
.activation(Activation.SOFTMAX)
.build(),
FREEZE_UNTIL_LAYER)
.build();
vgg16Transfer.setListeners(new ScoreIterationListener(5));
log.info("Modified VGG 16 Architecture for transfer learning");
log.info(vgg16Transfer.summary());
int iEpoch = 0;
int iIteration = 0;
while (iEpoch < EPOCH) {
while (neuralNetworkTrainingData.getTrainIterator().hasNext()) {
DataSet trainMiniBatchData = neuralNetworkTrainingData.getTrainIterator().next();
vgg16Transfer.fit(trainMiniBatchData);
saveProgressEveryConfiguredInterval(vgg16Transfer, iEpoch, iIteration);
iIteration++;
}
neuralNetworkTrainingData.getTrainIterator().reset();
iEpoch++;
evalOn(vgg16Transfer, neuralNetworkTrainingData.getTestIterator(), iEpoch);
}
}
Example #18
| Source File: TrainCifar10Model.java From Java-Machine-Learning-for-Computer-Vision with MIT License | 4 votes |
|
private void train() throws IOException {
ZooModel zooModel = VGG16.builder().build();
ComputationGraph vgg16 = (ComputationGraph) zooModel.initPretrained(PretrainedType.CIFAR10);
log.info(vgg16.summary());
IUpdater iUpdaterWithDefaultConfig = Updater.ADAM.getIUpdaterWithDefaultConfig();
iUpdaterWithDefaultConfig.setLrAndSchedule(0.1, null);
FineTuneConfiguration fineTuneConf = new FineTuneConfiguration.Builder()
.seed(1234)
// .weightInit(WeightInit.XAVIER)
.optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
.activation(Activation.RELU)
.updater(iUpdaterWithDefaultConfig)
.cudnnAlgoMode(ConvolutionLayer.AlgoMode.NO_WORKSPACE)
.miniBatch(true)
.inferenceWorkspaceMode(WorkspaceMode.ENABLED)
.trainingWorkspaceMode(WorkspaceMode.ENABLED)
.pretrain(true)
.backprop(true)
.build();
ComputationGraph cifar10 = new TransferLearning.GraphBuilder(vgg16)
.setWorkspaceMode(WorkspaceMode.ENABLED)
.fineTuneConfiguration(fineTuneConf)
.setInputTypes(InputType.convolutionalFlat(ImageUtils.HEIGHT,
ImageUtils.WIDTH, 3))
.removeVertexAndConnections("dense_2_loss")
.removeVertexAndConnections("dense_2")
.removeVertexAndConnections("dense_1")
.removeVertexAndConnections("dropout_1")
.removeVertexAndConnections("embeddings")
.removeVertexAndConnections("flatten_1")
.addLayer("dense_1", new DenseLayer.Builder()
.nIn(4096)
.nOut(EMBEDDINGS)
.activation(Activation.RELU).build(), "block3_pool")
.addVertex("embeddings", new L2NormalizeVertex(new int[]{}, 1e-12), "dense_1")
.addLayer("lossLayer", new CenterLossOutputLayer.Builder()
.lossFunction(LossFunctions.LossFunction.SQUARED_LOSS)
.activation(Activation.SOFTMAX).nIn(EMBEDDINGS).nOut(NUM_POSSIBLE_LABELS)
.lambda(LAMBDA).alpha(0.9)
.gradientNormalization(GradientNormalization.RenormalizeL2PerLayer).build(),
"embeddings")
.setOutputs("lossLayer")
.build();
log.info(cifar10.summary());
File rootDir = new File("CarTracking/train_from_video_" + NUM_POSSIBLE_LABELS);
DataSetIterator dataSetIterator = ImageUtils.createDataSetIterator(rootDir, NUM_POSSIBLE_LABELS, BATCH_SIZE);
DataSetIterator testIterator = ImageUtils.createDataSetIterator(rootDir, NUM_POSSIBLE_LABELS, BATCH_SIZE);
cifar10.setListeners(new ScoreIterationListener(2));
int iEpoch = I_EPOCH;
while (iEpoch < EPOCH_TRAINING) {
while (dataSetIterator.hasNext()) {
DataSet trainMiniBatchData = null;
try {
trainMiniBatchData = dataSetIterator.next();
} catch (Exception e) {
e.printStackTrace();
}
cifar10.fit(trainMiniBatchData);
}
iEpoch++;
String modelName = PREFIX + NUM_POSSIBLE_LABELS + "_epoch_data_e" + EMBEDDINGS + "_b" + BATCH_SIZE + "_" + iEpoch + ".zip";
saveProgress(cifar10, iEpoch, modelName);
testResults(cifar10, testIterator, iEpoch, modelName);
dataSetIterator.reset();
log.info("iEpoch = " + iEpoch);
}
}
Example #19
| Source File: KerasYolo9000PredictTest.java From deeplearning4j with Apache License 2.0 | 3 votes |
|
@Ignore
@Test
public void testYoloPredictionImport() throws Exception {
int HEIGHT = 416;
int WIDTH = 416;
INDArray indArray = Nd4j.create(HEIGHT, WIDTH, 3);
IMAGE_PREPROCESSING_SCALER.transform(indArray);
KerasLayer.registerCustomLayer("Lambda", KerasSpaceToDepth.class);
String h5_FILENAME = "modelimport/keras/examples/yolo/yolo-voc.h5";
ComputationGraph graph = KerasModelImport.importKerasModelAndWeights(h5_FILENAME, false);
double[][] priorBoxes = {{1.3221, 1.73145}, {3.19275, 4.00944}, {5.05587, 8.09892}, {9.47112, 4.84053}, {11.2364, 10.0071}};
INDArray priors = Nd4j.create(priorBoxes);
ComputationGraph model = new TransferLearning.GraphBuilder(graph)
.addLayer("outputs",
new org.deeplearning4j.nn.conf.layers.objdetect.Yolo2OutputLayer.Builder()
.boundingBoxPriors(priors)
.build(),
"conv2d_23")
.setOutputs("outputs")
.build();
ModelSerializer.writeModel(model, DL4J_MODEL_FILE_NAME, false);
ComputationGraph computationGraph = ModelSerializer.restoreComputationGraph(new File(DL4J_MODEL_FILE_NAME));
System.out.println(computationGraph.summary(InputType.convolutional(416, 416, 3)));
INDArray results = computationGraph.outputSingle(indArray);
}
