@Test
public void testRntn() throws Exception {
RNTN rntn =
new RNTN.Builder()
.setActivationFunction(Activations.tanh())
.setAdagradResetFrequency(1)
.setCombineClassification(true)
.setFeatureVectors(vec)
.setRandomFeatureVectors(false)
.setRng(new MersenneTwister(123))
.setUseTensors(true)
.build();
List<Tree> trees = vectorizer.getTreesWithLabels(sentence, Arrays.asList("LABEL"));
rntn.fit(trees);
}
/**
* Conf used for distributed deep learning
*
* @author Adam Gibson
*/
public class Conf implements Serializable, Cloneable {
private static final long serialVersionUID = 2994146097289344262L;
private Class<? extends BaseMultiLayerNetwork> multiLayerClazz;
private Class<? extends NeuralNetwork> neuralNetworkClazz;
private int k;
private long seed = 123;
private float corruptionLevel = 0.3f;
private float sparsity = 0;
private ActivationFunction function = Activations.sigmoid();
private ActivationFunction outputActivationFunction = Activations.softmax();
private int[] layerSizes = new int[] {300, 300, 300};
private int pretrainEpochs = 1000;
private int finetuneEpochs = 1000;
private float pretrainLearningRate = 0.01f;
private float finetuneLearningRate = 0.01f;
private int split = 10;
private int nIn = 1;
private int nOut = 1;
private int numPasses = 1;
private float momentum = 0.1f;
private boolean useRegularization = false;
private Object[] deepLearningParams;
private String masterUrl;
private float l2;
private Map<Integer, MatrixTransform> weightTransforms = new HashMap<>();
private Map<Integer, ActivationFunction> activationFunctionForLayer = new HashMap<>();
private Map<Integer, Integer> renderEpochsByLayer = new HashMap<>();
private int renderWeightEpochs = -1;
private String masterAbsPath;
private INDArray columnMeans;
private INDArray columnStds;
private boolean useAdaGrad = false;
private boolean useBackProp = true;
private float dropOut;
private OptimizationAlgorithm optimizationAlgorithm = OptimizationAlgorithm.CONJUGATE_GRADIENT;
private boolean normalizeZeroMeanAndUnitVariance;
private boolean scale;
private RBM.VisibleUnit visibleUnit = RBM.VisibleUnit.BINARY;
private RBM.HiddenUnit hiddenUnit = RBM.HiddenUnit.BINARY;
private String stateTrackerConnectionString;
private Map<Integer, RBM.VisibleUnit> visibleUnitByLayer = new HashMap<>();
private Map<Integer, RBM.HiddenUnit> hiddenUnitByLayer = new HashMap<>();
private Map<Integer, Float> learningRateForLayer = new HashMap<>();
private boolean roundCodeLayer = false;
private boolean normalizeCodeLayer = false;
private boolean lineSearchBackProp = false;
private boolean sampleHiddenActivations = false;
private Map<Integer, Boolean> sampleHiddenActivationsByLayer = new HashMap<>();
private boolean useDropConnect = false;
private float outputLayerDropOut = 0.0f;
private NeuralNetConfiguration conf;
public NeuralNetConfiguration getConf() {
return conf;
}
public void setConf(NeuralNetConfiguration conf) {
this.conf = conf;
}
public boolean isUseDropConnect() {
return useDropConnect;
}
public void setUseDropConnect(boolean useDropConnect) {
this.useDropConnect = useDropConnect;
}
public float getOutputLayerDropOut() {
return outputLayerDropOut;
}
public void setOutputLayerDropOut(float outputLayerDropOut) {
this.outputLayerDropOut = outputLayerDropOut;
}
public Map<Integer, Boolean> getSampleHiddenActivationsByLayer() {
return sampleHiddenActivationsByLayer;
}
public void setSampleHiddenActivationsByLayer(
Map<Integer, Boolean> sampleHiddenActivationsByLayer) {
this.sampleHiddenActivationsByLayer = sampleHiddenActivationsByLayer;
}
public boolean isSampleHiddenActivations() {
return sampleHiddenActivations;
}
public void setSampleHiddenActivations(boolean sampleHiddenActivations) {
this.sampleHiddenActivations = sampleHiddenActivations;
}
public boolean isLineSearchBackProp() {
return lineSearchBackProp;
}
public void setLineSearchBackProp(boolean lineSearchBackProp) {
this.lineSearchBackProp = lineSearchBackProp;
}
public Map<Integer, Integer> getRenderEpochsByLayer() {
return renderEpochsByLayer;
}
public void setRenderEpochsByLayer(Map<Integer, Integer> renderEpochsByLayer) {
this.renderEpochsByLayer = renderEpochsByLayer;
}
public boolean isNormalizeCodeLayer() {
return normalizeCodeLayer;
}
public void setNormalizeCodeLayer(boolean normalizeCodeLayer) {
this.normalizeCodeLayer = normalizeCodeLayer;
}
public boolean isRoundCodeLayer() {
return roundCodeLayer;
}
public void setRoundCodeLayer(boolean roundCodeLayer) {
this.roundCodeLayer = roundCodeLayer;
}
public Map<Integer, Float> getLearningRateForLayer() {
return learningRateForLayer;
}
public void setLearningRateForLayer(Map<Integer, Float> learningRateForLayer) {
this.learningRateForLayer = learningRateForLayer;
}
public ActivationFunction getOutputActivationFunction() {
return outputActivationFunction;
}
public void setOutputActivationFunction(ActivationFunction outputActivationFunction) {
this.outputActivationFunction = outputActivationFunction;
}
public Map<Integer, RBM.VisibleUnit> getVisibleUnitByLayer() {
return visibleUnitByLayer;
}
public void setVisibleUnitByLayer(Map<Integer, RBM.VisibleUnit> visibleUnitByLayer) {
this.visibleUnitByLayer = visibleUnitByLayer;
}
public Map<Integer, RBM.HiddenUnit> getHiddenUnitByLayer() {
return hiddenUnitByLayer;
}
public void setHiddenUnitByLayer(Map<Integer, RBM.HiddenUnit> hiddenUnitByLayer) {
this.hiddenUnitByLayer = hiddenUnitByLayer;
}
public Map<Integer, ActivationFunction> getActivationFunctionForLayer() {
return activationFunctionForLayer;
}
public void setActivationFunctionForLayer(
Map<Integer, ActivationFunction> activationFunctionForLayer) {
this.activationFunctionForLayer = activationFunctionForLayer;
}
public String getStateTrackerConnectionString() {
return stateTrackerConnectionString;
}
public void setStateTrackerConnectionString(String stateTrackerConnectionString) {
this.stateTrackerConnectionString = stateTrackerConnectionString;
}
public RBM.VisibleUnit getVisibleUnit() {
return visibleUnit;
}
public void setVisibleUnit(RBM.VisibleUnit visibleUnit) {
this.visibleUnit = visibleUnit;
}
public RBM.HiddenUnit getHiddenUnit() {
return hiddenUnit;
}
public void setHiddenUnit(RBM.HiddenUnit hiddenUnit) {
this.hiddenUnit = hiddenUnit;
}
public boolean isScale() {
return scale;
}
public void setScale(boolean scale) {
this.scale = scale;
}
public boolean isNormalizeZeroMeanAndUnitVariance() {
return normalizeZeroMeanAndUnitVariance;
}
public void setNormalizeZeroMeanAndUnitVariance(boolean normalizeZeroMeanAndUnitVariance) {
this.normalizeZeroMeanAndUnitVariance = normalizeZeroMeanAndUnitVariance;
}
public float getDropOut() {
return dropOut;
}
public void setDropOut(float dropOut) {
this.dropOut = dropOut;
}
/**
* Sets in and outs based on data
*
* @param data the data to use
*/
public void initFromData(DataSet data) {
setnIn(data.numInputs());
setnOut(data.numOutcomes());
}
public synchronized boolean isUseAdaGrad() {
return useAdaGrad;
}
public synchronized void setUseAdaGrad(boolean useAdaGrad) {
this.useAdaGrad = useAdaGrad;
}
public synchronized String getMasterAbsPath() {
return masterAbsPath;
}
public synchronized void setMasterAbsPath(String masterAbsPath) {
this.masterAbsPath = masterAbsPath;
}
public synchronized float getSparsity() {
return sparsity;
}
public synchronized void setSparsity(float sparsity) {
this.sparsity = sparsity;
}
public Map<Integer, MatrixTransform> getWeightTransforms() {
return weightTransforms;
}
public void setWeightTransforms(Map<Integer, MatrixTransform> weightTransforms) {
this.weightTransforms = weightTransforms;
}
public float getL2() {
return l2;
}
public void setL2(float l2) {
this.l2 = l2;
}
public String getMasterUrl() {
return masterUrl;
}
public void setMasterUrl(String masterUrl) {
this.masterUrl = masterUrl;
}
public float getMomentum() {
return momentum;
}
public void setMomentum(float momentum) {
this.momentum = momentum;
}
public boolean isUseRegularization() {
return useRegularization;
}
public void setUseRegularization(boolean useRegularization) {
this.useRegularization = useRegularization;
}
public Class<? extends BaseMultiLayerNetwork> getMultiLayerClazz() {
return multiLayerClazz;
}
public void setMultiLayerClazz(Class<? extends BaseMultiLayerNetwork> multiLayerClazz) {
this.multiLayerClazz = multiLayerClazz;
}
public Class<? extends NeuralNetwork> getNeuralNetworkClazz() {
return neuralNetworkClazz;
}
public void setNeuralNetworkClazz(Class<? extends NeuralNetwork> neuralNetworkClazz) {
this.neuralNetworkClazz = neuralNetworkClazz;
}
public int getK() {
return k;
}
public void setK(int k) {
this.k = k;
}
public long getSeed() {
return seed;
}
public void setSeed(long seed) {
this.seed = seed;
}
public float getCorruptionLevel() {
return corruptionLevel;
}
public void setCorruptionLevel(float corruptionLevel) {
this.corruptionLevel = corruptionLevel;
}
public ActivationFunction getFunction() {
return function;
}
public void setFunction(ActivationFunction function) {
this.function = function;
}
/**
* Returns the hidden layer sizes
*
* @return the hidden layer sizes applyTransformToDestination for this configuration
*/
public int[] getLayerSizes() {
return layerSizes;
}
/**
* Sets the hidden layer sizes
*
* @param layerSizes the layer sizes to use
*/
public void setLayerSizes(int[] layerSizes) {
this.layerSizes = layerSizes;
}
public void setLayerSizes(Integer[] layerSizes) {
this.layerSizes = new int[layerSizes.length];
for (int i = 0; i < layerSizes.length; i++) this.layerSizes[i] = layerSizes[i];
}
public int getPretrainEpochs() {
return pretrainEpochs;
}
public void setPretrainEpochs(int pretrainEpochs) {
this.pretrainEpochs = pretrainEpochs;
}
/**
* Sets the pretrain learning rate. Note that this will also be used for adagrad pretrain master
* learning rate
*
* @param pretrainLearningRate the learning rate to use
*/
public void setPretrainLearningRate(float pretrainLearningRate) {
this.pretrainLearningRate = pretrainLearningRate;
}
public float getFinetuneLearningRate() {
return finetuneLearningRate;
}
/**
* Sets the finetune learning rate. Note that this will also be used for adagrad finetune master
* learning rate
*
* @param finetuneLearningRate the learning rate to use
*/
public void setFinetuneLearningRate(float finetuneLearningRate) {
this.finetuneLearningRate = finetuneLearningRate;
}
public int getSplit() {
return split;
}
/**
* The optimal split is usually going to be something akin to the number of workers * the mini
* batch size.
*
* <p>Say if you have a system with 8 cores with 1 core per worker and a batch size of 10, you
* will want 80 as the batch size.
*
* @param split
*/
public void setSplit(int split) {
this.split = split;
}
public int getnIn() {
return nIn;
}
public void setnIn(int nIn) {
this.nIn = nIn;
}
public int getnOut() {
return nOut;
}
public void setnOut(int nOut) {
this.nOut = nOut;
}
/**
* Number of epochs to run
*
* @return the number of epochs to run
*/
public int getNumPasses() {
return numPasses;
}
public void setNumPasses(int numPasses) {
this.numPasses = numPasses;
}
public Object[] getDeepLearningParams() {
return deepLearningParams;
}
public void setDeepLearningParams(Object[] deepLearningParams) {
this.deepLearningParams = deepLearningParams;
}
public int getFinetuneEpochs() {
return finetuneEpochs;
}
public void setFinetuneEpochs(int finetuneEpochs) {
this.finetuneEpochs = finetuneEpochs;
}
public int getRenderWeightEpochs() {
return renderWeightEpochs;
}
public void setRenderWeightEpochs(int renderWeightEpochs) {
this.renderWeightEpochs = renderWeightEpochs;
}
public Conf copy() {
return SerializationUtils.clone(this);
}
public boolean isUseBackProp() {
return useBackProp;
}
public void setUseBackProp(boolean useBackProp) {
this.useBackProp = useBackProp;
}
/**
* Corruption level of 0.3 and learning rate of 0.01 and 1000 epochs
*
* @return
*/
public static Object[] getDefaultDenoisingAutoEncoderParams() {
return new Object[] {0.3, 0.01, 1000};
}
/**
* K of 1 and learning rate of 0.01 and 1000 epochs
*
* @return the default parameters for RBMs and DBNs
*/
public static Object[] getDefaultRbmParams() {
return new Object[] {1, 0.01, 1000};
}
public OptimizationAlgorithm getOptimizationAlgorithm() {
return optimizationAlgorithm;
}
public void setOptimizationAlgorithm(OptimizationAlgorithm optimizationAlgorithm) {
this.optimizationAlgorithm = optimizationAlgorithm;
}
/**
* Returns a multi layer network based on the configuration
*
* @return the initialized network
*/
public BaseMultiLayerNetwork init() {
if (getMultiLayerClazz().isAssignableFrom(DBN.class)) {
return new DBN.Builder()
.configure(conf)
.withClazz(getMultiLayerClazz())
.lineSearchBackProp(isLineSearchBackProp())
.hiddenLayerSizes(getLayerSizes())
.useDropConnection(isUseDropConnect())
.build();
} else {
return new BaseMultiLayerNetwork.Builder<>()
.withClazz(getMultiLayerClazz())
.hiddenLayerSizes(getLayerSizes())
.lineSearchBackProp(isLineSearchBackProp())
.useDropConnection(isUseDropConnect())
.build();
}
}
}