Shifu Plugin Mahout
protected void initModels() {
pmml = PMMLUtils.loadPMML(initPmmlPath);
List<MahoutDataPair> inputDataSet = new MahoutTestDataGenerator(inputData, pmml).getMahoutDataPair();
// create Mahout NeuralNetwork model
mlModel = new MultilayerPerceptron();
mlModel.addLayer(30, false, "Identity");// inputFields,isFinalLayer,squashFunction
mlModel.addLayer(45, false, "Sigmoid");
mlModel.addLayer(1, true, "Sigmoid");
for (MahoutDataPair pair : inputDataSet) {
mlModel.trainOnline(pair.getVectorAsInputVector());
}
}
protected void adaptModelToPMML() {
Model pmmlNN = pmml.getModels().get(0);
pmmlNN = new MahoutNeuralNetworkToPMMLModel().adaptMLModelToPMML(mlModel,(org.dmg.pmml.NeuralNetwork) pmmlNN);
pmml.getModels().set(0, pmmlNN);
}
private void evaluate(MahoutTestDataGenerator evalInput) {
for (Map<FieldName, String> map : evalInput.getEvaluatorInput()) {
double[] data = evalInput.normalizeData(context);
Assert.assertEquals(getPMMLEvaluatorResult(map),mlModel.getOutput(new DenseVector(data)).get(0), DELTA);
}
protected void initModels() {
pmml = PMMLUtils.loadPMML(initPmmlPath);
List<MahoutDataPair> inputDataSet = new MahoutTestDataGenerator(inputData, pmml).getMahoutDataPair();
lrModel = new OnlineLogisticRegression(2, 30, new L1());//numCategory, numFeatures, PriorFunction
for (MahoutDataPair pair : inputDataSet) {
lrModel.train(pair.getActual(), pair.getVector());
}
}
protected void adaptModelToPMML() {
Model pmmlLR = pmml.getModels().get(0);
pmmlLR = new MahoutLogisticRegressionToPMMLModel().adaptMLModelToPMML(mlModel,(org.dmg.pmml.RegressionModel) pmmlLR);
pmml.getModels().set(0, pmmlLR);
}
private void evaluate(MahoutTestDataGenerator evalInput) {
for (Map<FieldName, String> map : evalInput.getEvaluatorInput()) {
double[] data = evalInput.normalizeData(context);
Assert.assertEquals(getPMMLEvaluatorResult(map),lrModel.classifyScalar(new DenseVector(data)), DELTA);
}
1.In Mahout, the bias=1 comes as the first neuron in each layer that has bias neuron. Refer to the function trainByInstanceGradientDescent in the class org.apache.mahout.classifier.mlp.NeuralNetwork.
private Matrix[] trainByInstanceGradientDescent(Vector trainingInstance) {
int inputDimension = this.layerSizeList.get(0) - 1;
Vector inputInstance = new DenseVector(this.layerSizeList.get(0));
inputInstance.set(0, 1); // add bias
for (int i = 0; i < inputDimension; ++i) {
inputInstance.set(i + 1, trainingInstance.get(i));
}...
}
2.InputVector is the field vector followed by the actual vector. The len(inputVector) = len(fieldVector)+len(actual vector).
3.Since squashFunctionList in Mahout NeuralNetwork class is not accessible from outside, by default, we set squashFunction (ActivationFunction) to sigmoid.
Currently, Mahout only supports identity and sigmoid activation functions.
public static DoubleFunction getDerivativeDoubleFunction(String function) {
if (function.equalsIgnoreCase("Identity")) {
return derivativeIdentityFunction;
} else if (function.equalsIgnoreCase("Sigmoid")) {
return Functions.SIGMOIDGRADIENT;
} else {
throw new IllegalArgumentException("Function not supported.");
}
}
public class MahoutNeuralNetworkToPMML implements ModelToPMML<org.dmg.pmml.NeuralNetwork, org.apache.mahout.classifier.mlp.NeuralNetwork> {
public org.dmg.pmml.NeuralNetwork adaptMLModelToPMML(org.apache.mahout.classifier.mlp.NeuralNetwork nnModel, org.dmg.pmml.NeuralNetwork pmmlModel) {
pmmlModel = new NeuralNetworkModelIntegrator().adaptPMMLStatsModel(pmmlModel);
Matrix[] matrixList = nnModel.getWeightMatrices();
for (int layerIndex = 0; layerIndex < matrixList.length; layerIndex++) {
NeuralLayer layer = new NeuralLayer();
Matrix matrix = matrixList[layerIndex];
// TODO since squashFunctionList in Mahout is not accessible, by default, we set ActivationFunction to sigmoid
layer.setActivationFunction(ActivationFunctionType.LOGISTIC);
for (int j = 0; j < rowSize; j++) {
Neuron neuron = new Neuron(String.valueOf((layerIndex + 1) + "," + j));
neuron.setBias(0.0);// bias of each neuron, set to 0
for (int k = 1; k < columnSize; k++) {
neuron.withConnections(new Connection(String.valueOf(layerIndex + "," + (k - 1)), matrix.get(j,k)));
}
// bias neuron for each layer, set to bias=1
neuron.withConnections(new Connection(PluginConstants.biasValue, matrix.get(j, 0)));
}
layer.withNeurons(neuron);
}// finish build Neuron
pmmlModel.withNeuralLayers(layer);
}// finish build layer
pmmlModel.withNeuralOutputs(PMMLAdapterCommonUtil.getOutputFields(pmmlModel.getMiningSchema(), matrixList.length));
return pmmlModel;
}
}
public class MahoutLogisticRegressionToPMMLModel implements
ModelToPMML<org.dmg.pmml.RegressionModel, org.apache.mahout.classifier.sgd.OnlineLogisticRegression> {
public RegressionModel adaptMLModelToPMML(org.apache.mahout.classifier.sgd.OnlineLogisticRegression lrModel, org.dmg.pmml.RegressionModel pmmlModel) {
Matrix matrix = lrModel.getBeta();
int count = matrix.getNumNondefaultElements()[0];
double[] weights = new double[count];
for (int i = 0; i < count; i++)
weights[i] = matrix.get(0, i);
return PMMLAdapterCommonUtil.getRegressionTable(weights, 0, pmmlModel);
}
}
public class MahoutLogisticRegressionFromPMML implements PMMLToModel<OnlineLogisticRegression, RegressionModel> {
@Override
public OnlineLogisticRegression createMLModelFromPMML(RegressionModel pmmlModel) {
this.pmmlModel = pmmlModel;
initNNLayer();
setWeight();
return mlModel;
}
private void initModelStructure() {
mlModel = new OnlineLogisticRegression(2, numPredictors.size(),new L1());
}
private void setWeight() {
List<NumericPredictor> nPredictors = pmmlModel.getRegressionTables().get(0).getNumericPredictors();
for (int i = 0; i < nPredictors.size(); i++) {
mlModel.setBeta(0, i, nPredictors.get(i).getCoefficient());
}
}
The conversion from PMML to Mahout Neural Network is similar to the conversion to Mahout Logistic Regression Model
