multiclasswrapmh.java

Boosting算法软件包

		s[i] = l[i] * m_numLabels;
	    bags[0].subtractExampleList(s);
	    for (int j = 1; j < m_numLabels; j++) {
		for (i = 0; i < l.length; i++)
		    s[i]++;
		bags[j].subtractExampleList(s);
	    }
	}
    
	public void addBag(Bag b) {
	    MultiBag other = (MultiBag) b;
	    for (int j = 0; j < m_numLabels; j++)
		bags[j].addBag(other.bags[j]);
	}
    
	public void subtractBag(Bag b) {
	    MultiBag other = (MultiBag) b;
	    for (int j = 0; j < m_numLabels; j++)
		bags[j].subtractBag(other.bags[j]);
	}
    
	public void copyBag(Bag b) {
	    MultiBag other = (MultiBag) b;
	    for (int j = 0; j < m_numLabels; j++)
		bags[j].copyBag(other.bags[j]);
	}
    
	public void refresh(int index) {
	    int s = index * m_numLabels;
	    for (int j = 0; j < m_numLabels; j++)
		bags[j].refresh(s + j);
	}
    
	public void refreshList(int[] l) {
	    int i;
	    int[] s = new int[l.length];
	    for (i = 0; i < l.length; i++)
		s[i] = l[i] * m_numLabels;
	    bags[0].refreshList(s);
	    for (int j = 1; j < m_numLabels; j++) {
		for (i = 0; i < l.length; i++)
		    s[i]++;
		bags[j].refreshList(s);
	    }
	}
    
	/**
	 * Computes the loss for this bag.  This loss is only meaningful for
	 * additive losses.
	 */
	public double getLoss() {
	    double loss = 0.0;
	    for (int j = 0; j < m_numLabels; j++)
		loss += bags[j].getLoss();
	    return loss;
	}
    
	public double getLoss(int s)
	    throws jboost.NotSupportedException {
	    double loss = 0.0;
	    for (int j = 0; j < m_numLabels; j++)
		loss += bags[j].getLoss(s);
	    return loss;
	}
    }
  


    /**
     * This is the prediction class associated with this booster.
     * Each prediction is composed of an array of predictions from the
     * underlying booster, one for each class.  
     */
    class MultiPrediction extends Prediction {
	/**
	 * The predictions made.  Has same length as the number of classes.
	 */
	private Prediction[] preds;
    
	/**
	 * Constructor.
	 */
	private MultiPrediction() {
	    preds = new Prediction[m_numLabels];
	}
    
	public Object clone() {
	    MultiPrediction newpred = new MultiPrediction();
      
	    for (int j = 0; j < m_numLabels; j++) {
		newpred.preds[j] = (Prediction) preds[j].clone();
	    }
	    return newpred;
	}
    
	public Prediction add(Prediction p) {
	    for (int j = 0; j < m_numLabels; j++) {
		preds[j].add(((MultiPrediction) p).preds[j]);
	    }
	    return this;
	}
    
	public Prediction scale(double w) {
	    for (int j = 0; j < m_numLabels; j++)
		preds[j].scale(w);
	    return this;
	}
    
	public Prediction add(double w, Prediction p) {
	    for (int j = 0; j < m_numLabels; j++) {
		preds[j].add(w, ((MultiPrediction) p).preds[j]);
	    }
	    return this;
	}
    
	public double[] getMarginsSingleLabel(Label l) {
	    int maxClass = -1;
	    double maxScore = Double.MIN_VALUE;
	    int thisClass = -1;
	    double thisScore = 0;
	    for (int j = 0; j < m_numLabels; j++) {
		double predScore = preds[j].getClassScores()[1];
		if (l.getMultiValue(j)){
		    thisClass = j;
		    thisScore = predScore;
		} else if (maxScore < predScore) {
		    maxScore = predScore;
		    maxClass = j;
		}
	    }
	    double[] ret = new double[1];
	    ret[0] = thisScore - maxScore;
	    return ret;
	}
	
	public double[] getMarginsMultiLabel(Label l) {
	    double[] ret = new double[m_numLabels];
	    for (int j = 0; j < m_numLabels; j++) {
		ret[j] = preds[j].getMargins(new Label(l.getMultiValue(j) ?
						       1 : 0))[0];
	    }
	    return ret;
	}

	public double[] getMargins(Label l) {
	    if (m_isMultiLabel) 
		return getMarginsMultiLabel(l);
	    return getMarginsSingleLabel(l);
	}
    

	public double[] getClassScores() {
	    double[] scores = new double[m_numLabels];
	    double[] uscore;
      
	    for (int j = 0; j < m_numLabels; j++) {
		uscore = preds[j].getClassScores();
		scores[j] = uscore[1];
	    }
	    return scores;
	}
    
	/**
	 * Check to see if this MultiPrediction is the same as the other
	 * @param other
	 * @return true if all the Predictions of this object are equal to this other's
	 */
	public boolean equals(Prediction p) {
	    boolean retval= true;
	    MultiPrediction other= (MultiPrediction) p;
	    for (int k=0; k < m_numLabels; k++) {
		if (!preds[k].equals(other.preds[k])) {
		    retval= false;
		}
	    }
	    return retval;
	}
    
	public String toString() {
	    String s = "MultiPrediction.\n";
	    for (int j = 0; j < m_numLabels; j++)
		s += "prediction " + j + ": " + preds[j] + "\n";
	    return s;
	}
    
	public String shortText() {
	    String s = "[,"+preds[0];
	    for (int j = 0; j < m_numLabels; j++)
		s += ","+preds[j];
	    return s+"]";
	}
    
	public String cPreamble() {
	    String code = "";
      
	    code +=  "typedef double Prediction_t[" + m_numLabels + "];\n";
      
	    code +=  "#define reset_pred()  { \\\n";
	    for (int i = 0; i < m_numLabels; i++)
		code += "        p["+i+"] = 0.0; \\\n";
	    code += "     }\n";
      
	    code +=  "#define add_pred(";
	    for (int i = 0; i < m_numLabels; i++)
		code += (i == 0 ? "" : ",") + "X" + i;
	    code += ") { \\\n";
	    for (int i = 0; i < m_numLabels; i++)
		code += "        p["+i+"] += X"+i+"; \\\n";
	    code += "     }\n";
      
	    code += "#define finalize_pred() \\\n";
	    code += "        (r ? ( \\\n";
	    for (int i = 0; i < m_numLabels; i++)
		code += "                r["+i+"] = p["+i+"], \\\n";
	    code += "                p[0])       : p[0])\n";
      
	    return code;
	}
    
	public String javaPreamble() {
	    String code = "";
      
	    code += ""
		+ "  static private double[] p = new double[" + m_numLabels + "];\n"
		+ "  static private void reset_pred() {\n"
		+ "    java.util.Arrays.fill(p, 0.0);\n"
		+ "  }\n"
		+ "  static private void add_pred(";
	    for (int i = 0; i < m_numLabels; i++)
		code += (i == 0 ? "" : ",") + "double x" + i;
	    code += ") {\n";
	    for (int i = 0; i < m_numLabels; i++)
		code += "    p["+i+"] += x"+i+";\n";
	    code += ""
		+ "  }\n"
		+ "  static private double[] finalize_pred() {\n"
		+ "    return (double[]) p.clone();\n"
		+ "  }\n";
	    return code;
	}
    
	public double[] toCodeArray() {
	    return getClassScores();
	}
    }
  
    /** a main for testing */
    public static void main(String[] argv) {
	try {
	    AbstractBooster ada =
		new DebugWrap(new MulticlassWrapMH(new DebugWrap(new AdaBoost(0.0)), 2, true));
      
	    for(int i=0; i< 10; i++) {
		ada.addExample(i,new Label(i % 2));
	    }
      
	    ada.finalizeData();
	    if(Monitor.logLevel>3) Monitor.log(ada);
      
	    int[] elements = {1,2,6,3,4};
	    String s="\n Generating a bag with elements:\n" +elements[0];
	    for(int i=1; i<elements.length; i++) s+=", " + elements[i];
	    if(Monitor.logLevel>3) Monitor.log(s);
	    Bag bag = ada.newBag(elements);
	    if(Monitor.logLevel>3) Monitor.log(bag);
      
	    Prediction[] p;
	    p = ada.getPredictions(new Bag[] {bag}); // calc optimal prediction for this bag
	    if(Monitor.logLevel>3) Monitor.log("best prediction for this bag is " + 
					       p[0].getClassScores()[1]);
      
	    int[][] exampleList = new int[1][];
	    exampleList[0] = elements;
	    ada.update(p,exampleList);
	    if(Monitor.logLevel>3) Monitor.log("Adaboost after updating the m_weights is");
	    if(Monitor.logLevel>3) Monitor.log(ada);
	    bag.refreshList(elements);
	    if(Monitor.logLevel>3) Monitor.log("refreshed bag: "+bag);
	    if(Monitor.logLevel>3) Monitor.log("now best prediction for this bag is "+
					       (ada.getPredictions(new Bag[] {bag})[0]).getClassScores()[1]);
      
	}
	catch(Exception e) {
	    if(Monitor.logLevel>3) Monitor.log(e.getMessage());
	    e.printStackTrace();
	}
    
    
    }
  
    /* (non-Javadoc)
     * @see jboost.booster.Booster#init(jboost.controller.Configuration)
     */
    public void init(Configuration config) {
	// TODO Auto-generated method stub
    
    }
}