src/weka/classifiers/functions/MLPClassifier.java

/*

 *    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 2 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, write to the Free Software

 *    Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.

 */

/*

 *    MLPClassifier.java

 *    Copyright (C) 2012 University of Waikato, Hamilton, New Zealand

 */

package weka.classifiers.functions;

import java.util.Arrays;
import java.util.Collections;
import java.util.Enumeration;
import java.util.HashSet;
import java.util.Random;
import java.util.Set;
import java.util.Vector;
import java.util.concurrent.Callable;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.Future;

import weka.classifiers.Classifier;
import weka.classifiers.RandomizableClassifier;
import weka.core.Capabilities;
import weka.core.Capabilities.Capability;
import weka.core.ConjugateGradientOptimization;
import weka.core.Instance;
import weka.core.Instances;
import weka.core.Optimization;
import weka.core.Option;
import weka.core.RevisionUtils;
import weka.core.Utils;
import weka.core.WeightedInstancesHandler;
import weka.filters.Filter;
import weka.filters.unsupervised.attribute.NominalToBinary;
import weka.filters.unsupervised.attribute.RemoveUseless;
import weka.filters.unsupervised.attribute.ReplaceMissingValues;
import weka.filters.unsupervised.attribute.Standardize;

/**

 * <!-- globalinfo-start --> Trains a multilayer perceptron with one hidden

 * layer using WEKA's Optimization class by minimizing the squared error plus a

 * quadratic penalty with the BFGS method. Note that all attributes are

 * standardized. There are several parameters. The ridge parameter is used to

 * determine the penalty on the size of the weights. The number of hidden units

 * can also be specified. Note that large numbers produce long training times.

 * Finally, it is possible to use conjugate gradient descent rather than BFGS

 * updates, which may be faster for cases with many parameters. To improve

 * speed, an approximate version of the logistic function is used as the

 * activation function. 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 &lt;int&gt;

 *  Number of hidden units (default is 2).

 * </pre>

 * 

 * <pre>

 * -R &lt;double&gt;

 *  Ridge factor for quadratic penalty on weights (default is 0.01).

 * </pre>

 * 

 * <pre>

 * -O &lt;double&gt;

 *  Tolerance parameter for delta values (default is 1.0e-6).

 * </pre>

 * 

 * <pre>

 * -G

 *  Use conjugate gradient descent (recommended for many attributes).

 * </pre>

 * 

 * <pre>

 * -P &lt;int&gt;

 *  The size of the thread pool, for example, the number of cores in the CPU. (default 1)

 * </pre>

 * 

 * <pre>

 * -E &lt;int&gt;

 *  The number of threads to use, which should be &gt;= size of thread pool. (default 1)

 * </pre>

 * 

 * <pre>

 * -S &lt;num&gt;

 *  Random number seed.

 *  (default 1)

 * </pre>

 * 

 * <!-- options-end -->

 * 

 * @author Eibe Frank (eibe@cs.waikato.ac.nz)

 * @version $Revision: 10949 $

 */
public class MLPClassifier extends RandomizableClassifier implements WeightedInstancesHandler {

  /** For serialization */
  private static final long serialVersionUID = -3297474276438394644L;

  /**

   * Returns default capabilities of the classifier.

   * 

   * @return the capabilities of this classifier

   */
  @Override
  public Capabilities getCapabilities() {
    Capabilities result = super.getCapabilities();
    result.disableAll();

    // attributes
    result.enable(Capability.NOMINAL_ATTRIBUTES);
    result.enable(Capability.NUMERIC_ATTRIBUTES);
    result.enable(Capability.DATE_ATTRIBUTES);
    result.enable(Capability.MISSING_VALUES);

    // class
    result.enable(Capability.NOMINAL_CLASS);
    result.enable(Capability.MISSING_CLASS_VALUES);

    return result;
  }

  /**

   * Simple wrapper class needed to use the BFGS method implemented in

   * weka.core.Optimization.

   */
  protected class OptEng extends Optimization {

    /**

     * Returns the squared error given parameter values x.

     */
    @Override
    protected double objectiveFunction(double[] x) {

      m_MLPParameters = x;
      return calculateSE();
    }

    /**

     * Returns the gradient given parameter values x.

     */
    @Override
    protected double[] evaluateGradient(double[] x) {

      m_MLPParameters = x;
      return calculateGradient();
    }

    /**

     * The revision string.

     */
    @Override
    public String getRevision() {
      return RevisionUtils.extract("$Revision: 10949 $");
    }
  }

  /**

   * Simple wrapper class needed to use the CGD method implemented in

   * weka.core.ConjugateGradientOptimization.

   */
  protected class OptEngCGD extends ConjugateGradientOptimization {

    /**

     * Returns the squared error given parameter values x.

     */
    @Override
    protected double objectiveFunction(double[] x) {

      m_MLPParameters = x;
      return calculateSE();
    }

    /**

     * Returns the gradient given parameter values x.

     */
    @Override
    protected double[] evaluateGradient(double[] x) {

      m_MLPParameters = x;
      return calculateGradient();
    }

    /**

     * The revision string.

     */
    @Override
    public String getRevision() {
      return RevisionUtils.extract("$Revision: 10949 $");
    }
  }

  // The number of hidden units
  protected int m_numUnits = 2;

  // The class index of the dataset
  protected int m_classIndex = -1;

  // A reference to the actual data
  protected Instances m_data = null;

  // The number of classes in the data
  protected int m_numClasses = -1;

  // The number of attributes in the data
  protected int m_numAttributes = -1;

  // The parameter vector
  protected double[] m_MLPParameters = null;

  // Offset for output unit parameters
  protected int OFFSET_WEIGHTS = -1;

  // Offset for parameters of hidden units
  protected int OFFSET_ATTRIBUTE_WEIGHTS = -1;

  // The ridge parameter
  protected double m_ridge = 0.01;

  // Whether to use conjugate gradient descent rather than BFGS updates
  protected boolean m_useCGD = false;

  // Tolerance parameter for delta values
  protected double m_tolerance = 1.0e-6;

  // The number of threads to use to calculate gradient and squared error
  protected int m_numThreads = 1;

  // The size of the thread pool
  protected int m_poolSize = 1;

  // The standardization filer
  protected Filter m_Filter = null;

  // An attribute filter
  protected RemoveUseless m_AttFilter;

  // The filter used to make attributes numeric.
  protected NominalToBinary m_NominalToBinary;

  // The filter used to get rid of missing values.
  protected ReplaceMissingValues m_ReplaceMissingValues;

  // a ZeroR model in case no model can be built from the data
  protected Classifier m_ZeroR;

  // Thread pool
  protected transient ExecutorService m_Pool = null;

  /**

   * Method used to pre-process the data, perform clustering, and set the

   * initial parameter vector.

   */
  protected Instances initializeClassifier(Instances data) throws Exception {

    // can classifier handle the data?
    getCapabilities().testWithFail(data);

    data = new Instances(data);
    data.deleteWithMissingClass();

    // Make sure data is shuffled
    Random random = new Random(m_Seed);
    if (data.numInstances() > 1) {
      random = data.getRandomNumberGenerator(m_Seed);
    }
    data.randomize(random);

    // Replace missing values
    m_ReplaceMissingValues = new ReplaceMissingValues();
    m_ReplaceMissingValues.setInputFormat(data);
    data = Filter.useFilter(data, m_ReplaceMissingValues);

    // Remove useless attributes
    m_AttFilter = new RemoveUseless();
    m_AttFilter.setInputFormat(data);
    data = Filter.useFilter(data, m_AttFilter);

    // only class? -> build ZeroR model
    if (data.numAttributes() == 1) {
      System.err
        .println("Cannot build model (only class attribute present in data after removing useless attributes!), "
          + "using ZeroR model instead!");
      m_ZeroR = new weka.classifiers.rules.ZeroR();
      m_ZeroR.buildClassifier(data);
      return null;
    } else {
      m_ZeroR = null;
    }

    // Transform nominal attributes
    m_NominalToBinary = new NominalToBinary();
    m_NominalToBinary.setInputFormat(data);
    data = Filter.useFilter(data, m_NominalToBinary);

    // Standardize data
    m_Filter = new Standardize();
    m_Filter.setInputFormat(data);
    data = Filter.useFilter(data, m_Filter);

    m_classIndex = data.classIndex();
    m_numClasses = data.numClasses();
    m_numAttributes = data.numAttributes();

    // Set up array
    OFFSET_WEIGHTS = 0;
    OFFSET_ATTRIBUTE_WEIGHTS = (m_numUnits + 1) * m_numClasses;
    m_MLPParameters = new double[OFFSET_ATTRIBUTE_WEIGHTS + m_numUnits
      * m_numAttributes];

    // Initialize parameters
    for (int j = 0; j < m_numClasses; j++) {
      int offsetOW = OFFSET_WEIGHTS + (j * (m_numUnits + 1));
      for (int i = 0; i < m_numUnits; i++) {
        m_MLPParameters[offsetOW + i] = 0.1 * random.nextGaussian();
      }
      m_MLPParameters[offsetOW + m_numUnits] = 0.1 * random.nextGaussian();
    }
    for (int i = 0; i < m_numUnits; i++) {
      int offsetW = OFFSET_ATTRIBUTE_WEIGHTS + (i * m_numAttributes);
      for (int j = 0; j < m_numAttributes; j++) {
        m_MLPParameters[offsetW + j] = 0.1 * random.nextGaussian();
      }
    }

    return data;
  }

  /**

   * Builds the MLP network classifier based on the given dataset.

   */
  @Override
  public void buildClassifier(Instances data) throws Exception {

    // Set up the initial arrays
    m_data = initializeClassifier(data);
    if (m_data == null) {
      return;
    }

    // Initialise thread pool
    m_Pool = Executors.newFixedThreadPool(m_poolSize);

    // Apply optimization class to train the network
    Optimization opt = null;
    if (!m_useCGD) {
      opt = new OptEng();
    } else {
      opt = new OptEngCGD();
    }
    opt.setDebug(m_Debug);

    // No constraints
    double[][] b = new double[2][m_MLPParameters.length];
    for (int i = 0; i < 2; i++) {
      for (int j = 0; j < m_MLPParameters.length; j++) {
        b[i][j] = Double.NaN;
      }
    }

    m_MLPParameters = opt.findArgmin(m_MLPParameters, b);
    while (m_MLPParameters == null) {
      m_MLPParameters = opt.getVarbValues();
      if (m_Debug) {
        System.out.println("First set of iterations finished, not enough!");
      }
      m_MLPParameters = opt.findArgmin(m_MLPParameters, b);
    }
    if (m_Debug) {
      System.out.println("SE (normalized space) after optimization: "
        + opt.getMinFunction());
    }

    m_data = new Instances(m_data, 0); // Save memory

    // Shut down thread pool
    m_Pool.shutdown();
  }

  /**

   * Calculates the (penalized) squared error based on the current parameter

   * vector.

   */
  protected double calculateSE() {

    // Set up result set, and chunk size
    int chunksize = m_data.numInstances() / m_numThreads;
    Set<Future<Double>> results = new HashSet<Future<Double>>();

    // For each thread
    for (int j = 0; j < m_numThreads; j++) {

      // Determine batch to be processed
      final int lo = j * chunksize;
      final int hi = (j < m_numThreads - 1) ? (lo + chunksize) : m_data
        .numInstances();

      // Create and submit new job, where each instance in batch is processed
      Future<Double> futureSE = m_Pool.submit(new Callable<Double>() {
        @Override
        public Double call() {
          final double[] outputs = new double[m_numUnits];
          double SE = 0;
          for (int k = lo; k < hi; k++) {
            final Instance inst = m_data.instance(k);

            // Calculate necessary input/output values and error term
            calculateOutputs(inst, outputs, null);

            // 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 += inst.weight() * err * err;
            }
          }
          return SE;
        }
      });
      results.add(futureSE);
    }

    // Calculate SE
    double SE = 0;
    try {
      for (Future<Double> futureSE : results) {
        SE += futureSE.get();
      }
    } catch (Exception e) {
      System.out.println("Squared error could not be calculated.");
    }

    // Calculate sum of squared weights, excluding bias
    double squaredSumOfWeights = 0;
    for (int i = 0; i < m_numClasses; i++) {
      int offsetOW = OFFSET_WEIGHTS + (i * (m_numUnits + 1));
      for (int k = 0; k < m_numUnits; k++) {
        squaredSumOfWeights += m_MLPParameters[offsetOW + k]
          * m_MLPParameters[offsetOW + k];
      }
    }
    for (int k = 0; k < m_numUnits; k++) {
      int offsetW = OFFSET_ATTRIBUTE_WEIGHTS + k * m_numAttributes;
      for (int j = 0; j < m_classIndex; j++) {
        squaredSumOfWeights += m_MLPParameters[offsetW + j]
          * m_MLPParameters[offsetW + j];
      }
      for (int j = m_classIndex + 1; j < m_numAttributes; j++) {
        squaredSumOfWeights += m_MLPParameters[offsetW + j]
          * m_MLPParameters[offsetW + j];
      }
    }

    return ((m_ridge * squaredSumOfWeights) + (0.5 * SE))
      / m_data.sumOfWeights();
  }

  /**

   * Calculates the gradient based on the current parameter vector.

   */
  protected double[] calculateGradient() {

    // Set up result set, and chunk size
    int chunksize = m_data.numInstances() / m_numThreads;
    Set<Future<double[]>> results = new HashSet<Future<double[]>>();

    // For each thread
    for (int j = 0; j < m_numThreads; j++) {

      // Determine batch to be processed
      final int lo = j * chunksize;
      final int hi = (j < m_numThreads - 1) ? (lo + chunksize) : m_data
        .numInstances();

      // Create and submit new job, where each instance in batch is processed
      Future<double[]> futureGrad = m_Pool.submit(new Callable<double[]>() {
        @Override
        public double[] call() {

          final double[] outputs = new double[m_numUnits];
          final double[] deltaHidden = new double[m_numUnits];
          final double[] sigmoidDerivativeOutput = new double[1];
          final double[] sigmoidDerivativesHidden = new double[m_numUnits];
          final double[] localGrad = new double[m_MLPParameters.length];
          for (int k = lo; k < hi; k++) {
            final Instance inst = m_data.instance(k);
            calculateOutputs(inst, outputs, sigmoidDerivativesHidden);
            updateGradient(localGrad, inst, outputs, sigmoidDerivativeOutput,
              deltaHidden);
            updateGradientForHiddenUnits(localGrad, inst,
              sigmoidDerivativesHidden, deltaHidden);
          }
          return localGrad;
        }
      });
      results.add(futureGrad);
    }

    // Calculate final gradient
    double[] grad = new double[m_MLPParameters.length];
    try {
      for (Future<double[]> futureGrad : results) {
        double[] lg = futureGrad.get();
        for (int i = 0; i < lg.length; i++) {
          grad[i] += lg[i];
        }
      }
    } catch (Exception e) {
      System.out.println("Gradient could not be calculated.");
    }

    // For all network weights, perform weight decay
    for (int i = 0; i < m_numClasses; i++) {
      int offsetOW = OFFSET_WEIGHTS + (i * (m_numUnits + 1));
      for (int k = 0; k < m_numUnits; k++) {
        grad[offsetOW + k] += m_ridge * 2 * m_MLPParameters[offsetOW + k];
      }
    }
    for (int k = 0; k < m_numUnits; k++) {
      int offsetW = OFFSET_ATTRIBUTE_WEIGHTS + k * m_numAttributes;
      for (int j = 0; j < m_classIndex; j++) {
        grad[offsetW + j] += m_ridge * 2 * m_MLPParameters[offsetW + j];
      }
      for (int j = m_classIndex + 1; j < m_numAttributes; j++) {
        grad[offsetW + j] += m_ridge * 2 * m_MLPParameters[offsetW + j];
      }
    }

    double factor = 1.0 / m_data.sumOfWeights();
    for (int i = 0; i < grad.length; i++) {
      grad[i] *= factor;
    }

    return grad;
  }

  /**

   * 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 = inst.weight() * (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_MLPParameters[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;
    }
  }

  /**

   * Update the gradient for the weights in the hidden layer.

   */
  protected void updateGradientForHiddenUnits(double[] grad, Instance inst,

    double[] sigmoidDerivativesHidden, double[] deltaHidden) {

    // Finalize deltaHidden
    for (int i = 0; i < m_numUnits; i++) {
      deltaHidden[i] *= sigmoidDerivativesHidden[i];
    }

    // Update gradient for hidden units
    for (int i = 0; i < m_numUnits; i++) {

      // Skip calculations if update too small
      if (deltaHidden[i] <= m_tolerance && deltaHidden[i] >= -m_tolerance) {
        continue;
      }

      // Update gradient for all weights, including bias at classIndex
      int offsetW = OFFSET_ATTRIBUTE_WEIGHTS + i * m_numAttributes;
      for (int l = 0; l < m_classIndex; l++) {
        grad[offsetW + l] += deltaHidden[i] * inst.value(l);
      }
      grad[offsetW + m_classIndex] += deltaHidden[i];
      for (int l = m_classIndex + 1; l < m_numAttributes; l++) {
        grad[offsetW + l] += deltaHidden[i] * inst.value(l);
      }
    }
  }

  /**

   * Calculates the array of outputs of the hidden units. Also calculates

   * derivatives if d != null.

   */
  protected void calculateOutputs(Instance inst, double[] o, double[] d) {

    for (int i = 0; i < m_numUnits; i++) {
      int offsetW = OFFSET_ATTRIBUTE_WEIGHTS + i * m_numAttributes;
      double sum = 0;
      for (int j = 0; j < m_classIndex; j++) {
        sum += inst.value(j) * m_MLPParameters[offsetW + j];
      }
      sum += m_MLPParameters[offsetW + m_classIndex];
      for (int j = m_classIndex + 1; j < m_numAttributes; j++) {
        sum += inst.value(j) * m_MLPParameters[offsetW + j];
      }
      o[i] = sigmoid(-sum, d, i);
    }
  }

  /**

   * Calculates the output of output unit based on the given hidden layer

   * outputs. Also calculates the derivative if d != null.

   */
  protected double getOutput(int unit, double[] outputs, double[] d) {

    int offsetOW = OFFSET_WEIGHTS + (unit * (m_numUnits + 1));
    double result = 0;
    for (int i = 0; i < m_numUnits; i++) {
      result += m_MLPParameters[offsetOW + i] * outputs[i];
    }
    result += m_MLPParameters[offsetOW + m_numUnits];
    return sigmoid(-result, d, 0);
  }

  /**

   * 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;
  }

  /**

   * Calculates the output of the network after the instance has been piped

   * through the fliters to replace missing values, etc.

   */
  @Override
  public double[] distributionForInstance(Instance inst) throws Exception {

    m_ReplaceMissingValues.input(inst);
    inst = m_ReplaceMissingValues.output();
    m_AttFilter.input(inst);
    inst = m_AttFilter.output();

    // default model?
    if (m_ZeroR != null) {
      return m_ZeroR.distributionForInstance(inst);
    }

    m_NominalToBinary.input(inst);
    inst = m_NominalToBinary.output();
    m_Filter.input(inst);
    inst = m_Filter.output();

    double[] dist = new double[m_numClasses];
    double[] outputs = new double[m_numUnits];
    calculateOutputs(inst, outputs, null);
    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;
  }

  /**

   * This will return a string describing the classifier.

   * 

   * @return The string.

   */
  public String globalInfo() {

    return "Trains a multilayer perceptron with one hidden layer using WEKA's Optimization class"
      + " by minimizing the squared error plus a quadratic penalty with the BFGS method."
      + " Note that all attributes are standardized. There are several parameters. The"
      + " ridge parameter is used to determine the penalty on the size of the weights. The"
      + " number of hidden units can also be specified. Note that large"
      + " numbers produce long training times. Finally, it is possible to use conjugate gradient"
      + " descent rather than BFGS updates, which may be faster for cases with many parameters."
      + " To improve speed, an approximate version of the logistic function is used as the"
      + " activation function. 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.";
  }

  /**

   * @return a string to describe the option

   */
  public String toleranceTipText() {

    return "The tolerance parameter for the delta values.";
  }

  /**

   * Gets the tolerance parameter for the delta values.

   */
  public double getTolerance() {

    return m_tolerance;
  }

  /**

   * Sets the tolerance parameter for the delta values.

   */
  public void setTolerance(double newTolerance) {

    m_tolerance = newTolerance;
  }

  /**

   * @return a string to describe the option

   */
  public String numFunctionsTipText() {

    return "The number of hidden units to use.";
  }

  /**

   * Gets the number of functions.

   */
  public int getNumFunctions() {

    return m_numUnits;
  }

  /**

   * Sets the number of functions.

   */
  public void setNumFunctions(int newNumFunctions) {

    m_numUnits = newNumFunctions;
  }

  /**

   * @return a string to describe the option

   */
  public String ridgeTipText() {

    return "The ridge penalty factor for the quadratic penalty on the weights.";
  }

  /**

   * Gets the value of the ridge parameter.

   */
  public double getRidge() {

    return m_ridge;
  }

  /**

   * Sets the value of the ridge parameter.

   */
  public void setRidge(double newRidge) {

    m_ridge = newRidge;
  }

  /**

   * @return a string to describe the option

   */
  public String useCGDTipText() {

    return "Whether to use conjugate gradient descent (potentially useful for many parameters).";
  }

  /**

   * Gets whether to use CGD.

   */
  public boolean getUseCGD() {

    return m_useCGD;
  }

  /**

   * Sets whether to use CGD.

   */
  public void setUseCGD(boolean newUseCGD) {

    m_useCGD = newUseCGD;
  }

  /**

   * @return a string to describe the option

   */
  public String numThreadsTipText() {

    return "The number of threads to use, which should be >= size of thread pool.";
  }

  /**

   * Gets the number of threads.

   */
  public int getNumThreads() {

    return m_numThreads;
  }

  /**

   * Sets the number of threads

   */
  public void setNumThreads(int nT) {

    m_numThreads = nT;
  }

  /**

   * @return a string to describe the option

   */
  public String poolSizeTipText() {

    return "The size of the thread pool, for example, the number of cores in the CPU.";
  }

  /**

   * Gets the number of threads.

   */
  public int getPoolSize() {

    return m_poolSize;
  }

  /**

   * Sets the number of threads

   */
  public void setPoolSize(int nT) {

    m_poolSize = nT;
  }

  /**

   * Returns an enumeration describing the available options.

   * 

   * @return an enumeration of all the available options.

   */
  @Override
  public Enumeration<Option> listOptions() {

    Vector<Option> newVector = new Vector<Option>(6);

    newVector.addElement(new Option(
      "\tNumber of hidden units (default is 2).\n", "N", 1, "-N <int>"));

    newVector.addElement(new Option(
      "\tRidge factor for quadratic penalty on weights (default is 0.01).\n",
      "R", 1, "-R <double>"));
    newVector.addElement(new Option(
      "\tTolerance parameter for delta values (default is 1.0e-6).\n", "O", 1,
      "-O <double>"));
    newVector.addElement(new Option(
      "\tUse conjugate gradient descent (recommended for many attributes).\n",
      "G", 0, "-G"));
    newVector.addElement(new Option(
      "\t" + poolSizeTipText() + " (default 1)\n", "P", 1, "-P <int>"));
    newVector.addElement(new Option("\t" + numThreadsTipText()
      + " (default 1)\n", "E", 1, "-E <int>"));

    newVector.addAll(Collections.list(super.listOptions()));

    return newVector.elements();
  }

  /**

   * Parses a given list of options.

   * <p/>

   * 

   * <!-- options-start --> Valid options are:

   * <p/>

   * 

   * <pre>

   * -N &lt;int&gt;

   *  Number of hidden units (default is 2).

   * </pre>

   * 

   * <pre>

   * -R &lt;double&gt;

   *  Ridge factor for quadratic penalty on weights (default is 0.01).

   * </pre>

   * 

   * <pre>

   * -O &lt;double&gt;

   *  Tolerance parameter for delta values (default is 1.0e-6).

   * </pre>

   * 

   * <pre>

   * -G

   *  Use conjugate gradient descent (recommended for many attributes).

   * </pre>

   * 

   * <pre>

   * -P &lt;int&gt;

   *  The size of the thread pool, for example, the number of cores in the CPU. (default 1)

   * </pre>

   * 

   * <pre>

   * -E &lt;int&gt;

   *  The number of threads to use, which should be &gt;= size of thread pool. (default 1)

   * </pre>

   * 

   * <pre>

   * -S &lt;num&gt;

   *  Random number seed.

   *  (default 1)

   * </pre>

   * 

   * <!-- options-end -->

   * 

   * Options after -- are passed to the designated classifier.

   * <p>

   * 

   * @param options the list of options as an array of strings

   * @throws Exception if an option is not supported

   */
  @Override
  public void setOptions(String[] options) throws Exception {

    String numFunctions = Utils.getOption('N', options);
    if (numFunctions.length() != 0) {
      setNumFunctions(Integer.parseInt(numFunctions));
    } else {
      setNumFunctions(2);
    }
    String Ridge = Utils.getOption('R', options);
    if (Ridge.length() != 0) {
      setRidge(Double.parseDouble(Ridge));
    } else {
      setRidge(0.01);
    }
    String Tolerance = Utils.getOption('O', options);
    if (Tolerance.length() != 0) {
      setTolerance(Double.parseDouble(Tolerance));
    } else {
      setTolerance(1.0e-6);
    }
    m_useCGD = Utils.getFlag('G', options);
    String PoolSize = Utils.getOption('P', options);
    if (PoolSize.length() != 0) {
      setPoolSize(Integer.parseInt(PoolSize));
    } else {
      setPoolSize(1);
    }
    String NumThreads = Utils.getOption('E', options);
    if (NumThreads.length() != 0) {
      setNumThreads(Integer.parseInt(NumThreads));
    } else {
      setNumThreads(1);
    }

    super.setOptions(options);

    Utils.checkForRemainingOptions(options);
  }

  /**

   * Gets the current settings of the Classifier.

   * 

   * @return an array of strings suitable for passing to setOptions

   */
  @Override
  public String[] getOptions() {

    Vector<String> options = new Vector<String>();

    options.add("-N");
    options.add("" + getNumFunctions());

    options.add("-R");
    options.add("" + getRidge());

    options.add("-O");
    options.add("" + getTolerance());

    if (m_useCGD) {
      options.add("-G");
    }

    options.add("-P");
    options.add("" + getPoolSize());

    options.add("-E");
    options.add("" + getNumThreads());

    Collections.addAll(options, super.getOptions());

    return options.toArray(new String[0]);
  }

  /**

   * Outputs the network as a string.

   */
  @Override
  public String toString() {

    if (m_ZeroR != null) {
      return m_ZeroR.toString();
    }

    if (m_MLPParameters == null) {
      return "Classifier not built yet.";
    }

    String s = "MLPClassifier with ridge value " + getRidge() + " and "
      + getNumFunctions() + " hidden units (useCGD=" + getUseCGD() + ")\n\n";

    for (int i = 0; i < m_numUnits; i++) {
      for (int j = 0; j < m_numClasses; j++) {
        s += "Output unit " + j + " weight for hidden unit " + i + ": "
          + m_MLPParameters[OFFSET_WEIGHTS + j * (m_numUnits + 1) + i] + "\n";
      }
      s += "\nHidden unit " + i + " weights:\n\n";
      for (int j = 0; j < m_numAttributes; j++) {
        if (j != m_classIndex) {
          s += m_MLPParameters[OFFSET_ATTRIBUTE_WEIGHTS + (i * m_numAttributes)
            + j]
            + " " + m_data.attribute(j).name() + "\n";
        }
      }
      s += "\nHidden unit "
        + i
        + " bias: "
        + m_MLPParameters[OFFSET_ATTRIBUTE_WEIGHTS
          + (i * m_numAttributes + m_classIndex)] + "\n\n";
    }
    for (int j = 0; j < m_numClasses; j++) {
      s += "Output unit " + j + " bias: "
        + m_MLPParameters[OFFSET_WEIGHTS + j * (m_numUnits + 1) + m_numUnits]
        + "\n";
    }

    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 MLPClassifier(), argv);
  }
}