Java tutorial
/* * This program is free software: you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation, either version 3 of the License, or * (at your option) any later version. * * This program is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * GNU General Public License for more details. * * You should have received a copy of the GNU General Public License * along with this program. If not, see <http://www.gnu.org/licenses/>. */ /* * RBFClassifier.java * Copyright (C) 2012 University of Waikato, Hamilton, New Zealand */ package br.com.ufu.lsi.rebfnetwork; import weka.core.Instance; import weka.core.Utils; import weka.core.Capabilities; import weka.core.Capabilities.Capability; import weka.classifiers.Classifier; import weka.classifiers.Evaluation; import java.util.Random; import java.util.Arrays; /** <!-- globalinfo-start --> * Class implementing radial basis function networks for classification, trained in a fully supervised manner using WEKA's Optimization class by minimizing squared error with the BFGS method. Note that all attributes are normalized into the [0,1] scale. The initial centers for the Gaussian radial basis functions are found using WEKA's SimpleKMeans. The initial sigma values are set to the maximum distance between any center and its nearest neighbour in the set of centers. There are several parameters. The ridge parameter is used to penalize the size of the weights in the output layer. The number of basis functions can also be specified. Note that large numbers produce long training times. Another option determines whether one global sigma value is used for all units (fastest), whether one value is used per unit (common practice, it seems, and set as the default), or a different value is learned for every unit/attribute combination. It is also possible to learn attribute weights for the distance function. (The square of the value shown in the output is used.) Finally, it is possible to use conjugate gradient descent rather than BFGS updates, which can be faster for cases with many parameters, and to use normalized basis functions instead of unnormalized ones. To improve speed, an approximate version of the logistic function is used as the activation function in the output layer. Also, if delta values in the backpropagation step are within the user-specified tolerance, the gradient is not updated for that particular instance, which saves some additional time. Paralled calculation of squared error and gradient is possible when multiple CPU cores are present. Data is split into batches and processed in separate threads in this case. Note that this only improves runtime for larger datasets. Nominal attributes are processed using the unsupervised NominalToBinary filter and missing values are replaced globally using ReplaceMissingValues. * <p/> <!-- globalinfo-end --> * <!-- options-start --> * Valid options are: <p/> * * <pre> -N <int> * Number of Gaussian basis functions (default is 2). * </pre> * * <pre> -R <double> * Ridge factor for quadratic penalty on output weights (default is 0.01). * </pre> * * <pre> -L <double> * Tolerance parameter for delta values (default is 1.0e-6). * </pre> * * <pre> -C <1|2|3> * The scale optimization option: global scale (1), one scale per unit (2), scale per unit and attribute (3) (default is 2). * </pre> * * <pre> -G * Use conjugate gradient descent (recommended for many attributes). * </pre> * * <pre> -O * Use normalized basis functions. * </pre> * * <pre> -A * Use attribute weights. * </pre> * * <pre> -P <int> * The size of the thread pool, for example, the number of cores in the CPU. (default 1) * </pre> * * <pre> -E <int> * The number of threads to use, which should be >= size of thread pool. (default 1) * </pre> * * <pre> -S <num> * Random number seed. * (default 1)</pre> * * <pre> -D * If set, classifier is run in debug mode and * may output additional info to the console</pre> * <!-- options-end --> * * @author Eibe Frank (eibe@cs.waikato.ac.nz) * @version $Revision: 9402 $ */ public class RBFClassifier extends RBFModel { /** For serialization */ private static final long serialVersionUID = -7847475556438394611L; /** * Returns default capabilities of the classifier. * * @return the capabilities of this classifier */ public Capabilities getCapabilities() { Capabilities result = super.getCapabilities(); // class result.enable(Capability.NOMINAL_CLASS); result.enable(Capability.MISSING_CLASS_VALUES); return result; } /** * Initialise output layer. */ protected void initializeOutputLayer(Random random) { for (int i = 0; i < m_numUnits + 1; i++) { for (int j = 0; j < m_numClasses; j++) { m_RBFParameters[OFFSET_WEIGHTS + (j * (m_numUnits + 1)) + i] = 0.1 * random.nextGaussian(); } } } /** * Calculates error for single instance. */ protected double calculateError(double[] outputs, Instance inst) { // Want to calculate squared error double SE = 0; // For all class values for (int i = 0; i < m_numClasses; i++) { // Get target (make them slightly different from 0/1 for better convergence) final double target = ((int) inst.value(m_classIndex) == i) ? 0.99 : 0.01; // Add to squared error final double err = getOutput(i, outputs, null) - target; SE += err * err; } return SE; } /** * Postprocess squared error if desired. Returns argument unmodified by default. */ protected double postprocessError(double error) { // Calculate squared sum of weights double squaredSumOfWeights = 0; for (int k = 0; k < m_numUnits; k++) { for (int i = 0; i < m_numClasses; i++) { squaredSumOfWeights += m_RBFParameters[OFFSET_WEIGHTS + (i * (m_numUnits + 1)) + k] * m_RBFParameters[OFFSET_WEIGHTS + (i * (m_numUnits + 1)) + k]; } } return (error + m_ridge * squaredSumOfWeights) / m_data.numInstances(); } /** * Postprocess gradient if desired. */ protected void postprocessGradient(double[] grad) { // For each output weight, include effect of ridge for (int k = 0; k < m_numUnits; k++) { for (int i = 0; i < m_numClasses; i++) { grad[OFFSET_WEIGHTS + (i * (m_numUnits + 1)) + k] += m_ridge * 2 * m_RBFParameters[OFFSET_WEIGHTS + (i * (m_numUnits + 1)) + k]; } } double factor = 1.0 / m_data.numInstances(); for (int i = 0; i < grad.length; i++) { grad[i] *= factor; } } /** * Update the gradient for the weights in the output layer. */ protected void updateGradient(double[] grad, Instance inst, double[] outputs, double[] sigmoidDerivativeOutput, double[] deltaHidden) { // Initialise deltaHidden Arrays.fill(deltaHidden, 0.0); // For all output units for (int j = 0; j < m_numClasses; j++) { // Get output from output unit j double pred = getOutput(j, outputs, sigmoidDerivativeOutput); // Get target (make them slightly different from 0/1 for better convergence) double target = ((int) inst.value(m_classIndex) == j) ? 0.99 : 0.01; // Calculate delta from output unit double deltaOut = (pred - target) * sigmoidDerivativeOutput[0]; // Go to next output unit if update too small if (deltaOut <= m_tolerance && deltaOut >= -m_tolerance) { continue; } // Establish offset int offsetOW = OFFSET_WEIGHTS + (j * (m_numUnits + 1)); // Update deltaHidden for (int i = 0; i < m_numUnits; i++) { deltaHidden[i] += deltaOut * m_RBFParameters[offsetOW + i]; } // Update gradient for output weights for (int i = 0; i < m_numUnits; i++) { grad[offsetOW + i] += deltaOut * outputs[i]; } // Update gradient for bias grad[offsetOW + m_numUnits] += deltaOut; } } /** * Calculates the output of the network based on the given * hidden layer outputs. Also calculates the derivative * if d != null. */ protected double getOutput(int unit, double[] outputs, double[] d) { double result = 0; for (int i = 0; i < m_numUnits; i++) { result += m_RBFParameters[OFFSET_WEIGHTS + (unit * (m_numUnits + 1)) + i] * outputs[i]; } result += m_RBFParameters[OFFSET_WEIGHTS + (unit * (m_numUnits + 1)) + m_numUnits]; return sigmoid(-result, d, 0); } /** * Computes approximate exponential function. Derivative is * stored in second argument at given index if d != null. */ protected double approxExp(double x, double[] d, int index) { double y = 1.0 + x / 4096.0; x = y * y; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; // Compute derivative if desired if (d != null) { d[index] = x / y; } return x; } /** * Computes approximate sigmoid function. Derivative is * stored in second argument at given index if d != null. */ protected double sigmoid(double x, double[] d, int index) { // Compute approximate sigmoid double y = 1.0 + x / 4096.0; x = y * y; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; x *= x; double output = 1.0 / (1.0 + x); // Compute derivative if desired if (d != null) { d[index] = output * (1.0 - output) / y; } return output; } /** * Gets output "distribution" based on hidden layer outputs. */ protected double[] getDistribution(double[] outputs) { double[] dist = new double[m_numClasses]; for (int i = 0; i < m_numClasses; i++) { dist[i] = getOutput(i, outputs, null); if (dist[i] < 0) { dist[i] = 0; } else if (dist[i] > 1) { dist[i] = 1; } } Utils.normalize(dist); return dist; } /** * Outputs the network as a string. */ public String toString() { if (m_RBFParameters == null) { return "Classifier not built yet."; } String s = ""; for (int i = 0; i < m_numUnits; i++) { if (i > 0) { s += "\n\n"; } s += "Output weights for different classes:\n"; for (int j = 0; j < m_numClasses; j++) { s += m_RBFParameters[OFFSET_WEIGHTS + (j * (m_numUnits + 1)) + i] + "\t"; } s += "\n\nUnit center:\n"; for (int j = 0; j < m_numAttributes; j++) { if (j != m_classIndex) { s += m_RBFParameters[OFFSET_CENTERS + (i * m_numAttributes) + j] + "\t"; } } if (m_scaleOptimizationOption == USE_SCALE_PER_UNIT_AND_ATTRIBUTE) { s += "\n\nUnit scales:\n"; for (int j = 0; j < m_numAttributes; j++) { if (j != m_classIndex) { s += m_RBFParameters[OFFSET_SCALES + (i * m_numAttributes) + j] + "\t"; } } } else if (m_scaleOptimizationOption == USE_SCALE_PER_UNIT) { s += "\n\nUnit scale:\n"; s += m_RBFParameters[OFFSET_SCALES + i] + "\t"; } } if (m_scaleOptimizationOption == USE_GLOBAL_SCALE) { s += "\n\nScale:\n"; s += m_RBFParameters[OFFSET_SCALES] + "\t"; } if (m_useAttributeWeights) { s += "\n\nAttribute weights:\n"; for (int j = 0; j < m_numAttributes; j++) { if (j != m_classIndex) { s += m_RBFParameters[OFFSET_ATTRIBUTE_WEIGHTS + j] + "\t"; } } } s += "\n\nBias weights for different classes:\n"; for (int j = 0; j < m_numClasses; j++) { s += m_RBFParameters[OFFSET_WEIGHTS + (j * (m_numUnits + 1)) + m_numUnits] + "\t"; } return s; } /** * Main method to run the code from the command-line using * the standard WEKA options. */ public static void main(String[] argv) { runClassifier(new RBFClassifier(), argv); } }