import java.util.Date;
import java.util.HashSet;
import java.util.Iterator;
import java.util.Random;

import jnisvmlight.LabeledFeatureVector;
import jnisvmlight.SVMLightModel;
import jnisvmlight.SVMLightInterface;
import jnisvmlight.TrainingParameters;

public class JNI_SVMLight_Test {

  public static int N = 2000; // number of training docs

  public static int M = 100; // max. number of features per doc

  public static void main(String[] args) throws Exception {
    Random rd = new java.util.Random(new Date().getTime());

    // The trainer interface with the native communication to the SVM-light shared
    // libraries
    SVMLightInterface trainer = new SVMLightInterface();

    // The training data
    LabeledFeatureVector[] traindata = new LabeledFeatureVector[N];

    // Sort all feature vectors in ascedending order of feature dimensions
    // before training the model
    SVMLightInterface.SORT_INPUT_VECTORS = true;

    // Generate a randomly initialized array of N sparse training vectors
    // with at most M randomly chosen feature dimensions and values
    System.out.print("GENERATING A RANDOM TRAINING SET [N=" + N + ", M=" + M
        + "] ..");
    for (int i = 0; i < N; i++) {

      int nDims = 1 + rd.nextInt(M - 1);
      int[] dims = new int[nDims];
      double[] values = new double[nDims];

      // Generate a number of distinct random feature dimensions
      HashSet hashedDims = new HashSet();
      while (hashedDims.size() < nDims) {
        hashedDims.add(new Integer(1 + rd.nextInt(M - 1)));
      }

      // Fill the vectors
      int j = 0;
      for (Iterator iterator = hashedDims.iterator(); iterator.hasNext(); j++) {
        dims[j] = ((Integer) iterator.next()).intValue();
        values[j] = rd.nextDouble();
      }
      if (i % 10 == 0) {
        System.out.print(i + ".");
      }

      // Store dimension/value pairs in new LabeledFeatureVector object
      traindata[i] = new LabeledFeatureVector(((i < (N / 2)) ? +1 : -1), dims,
          values);

      // Use cosine similarities (LinearKernel with L2-normalized input vectors)
      traindata[i].normalizeL2();
    }
    System.out.println(" DONE.");

    // Initialize a new TrainingParamteres object with the default SVM-light
    // values
    TrainingParameters tp = new TrainingParameters();

    // Switch on some debugging output
    tp.getLearningParameters().verbosity = 1;

    System.out.println("\nTRAINING SVM-light MODEL ..");
    SVMLightModel model = trainer.trainModel(traindata, tp);
    System.out.println(" DONE.");

    // Use this to store a model to a file or read a model from a URL.
    //model.writeModelToFile("jni_model.dat");
    //model = SVMLightModel.readSVMLightModelFromURL(new java.io.File("jni_model.dat").toURL());
    
    // Use the classifier on the randomly created feature vectors
    System.out.println("\nVALIDATING SVM-light MODEL in Java..");
    int precision = 0;
    for (int i = 0; i < N; i++) {

      // Classify a test vector using the Java object
      // (in a real application, this should not be one of the training vectors)
      double d = model.classify(traindata[i]);
      if ((traindata[i].getLabel() < 0 && d < 0)
          || (traindata[i].getLabel() > 0 && d > 0)) {
        precision++;
      }
      if (i % 10 == 0) {
        System.out.print(i + ".");
      }
    }
    System.out.println(" DONE.");
    System.out.println("\n" + ((double) precision / N)
        + " PRECISION=RECALL ON RANDOM TRAINING SET.");

    System.out.println("\nVALIDATING SVM-light MODEL in Native Mode..");
    precision = 0;
    for (int i = 0; i < N; i++) {

      // Classify a test vector using the Native Interface
      // (in a real application, this should not be one of the training vectors)
      double d = trainer.classifyNative(traindata[i]);
      if ((traindata[i].getLabel() < 0 && d < 0)
          || (traindata[i].getLabel() > 0 && d > 0)) {
        precision++;
      }
      if (i % 10 == 0) {
        System.out.print(i + ".");
      }
    }
    System.out.println(" DONE.");
    System.out.println("\n" + ((double) precision / N)
        + " PRECISION=RECALL ON RANDOM TRAINING SET.");

    // An alternative way to invoke the SVM-light interface similar to using the
    // command line parameters.
    //System.out
    //    .println("\nTRAINING SVM-light MODEL WITH COMMAND LINE PARAMETERS ..");
    //model = trainer.trainModel(traindata, new String[] { "-z", "c" });

    System.out.println(" DONE.");
  }
}