validationchain.java

一个很好的LIBSVM的JAVA源码。对于要研究和改进SVM算法的学者。可以参考。来自数据挖掘工具YALE工具包。

/*
 *  YALE - Yet Another Learning Environment
 *  Copyright (C) 2001-2004
 *      Simon Fischer, Ralf Klinkenberg, Ingo Mierswa, 
 *          Katharina Morik, Oliver Ritthoff
 *      Artificial Intelligence Unit
 *      Computer Science Department
 *      University of Dortmund
 *      44221 Dortmund,  Germany
 *  email: [e-mail]
 *  web:   http://yale.cs.uni-dortmund.de/
 *
 *  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., 59 Temple Place, Suite 330, Boston, MA 02111-1307
 *  USA.
 */
package edu.udo.cs.yale.operator.validation;

import edu.udo.cs.yale.tools.math.AverageVector;
import edu.udo.cs.yale.example.ExampleSet;
import edu.udo.cs.yale.example.Attribute;
import edu.udo.cs.yale.operator.Operator;
import edu.udo.cs.yale.operator.OperatorChain;
import edu.udo.cs.yale.operator.IOContainer;
import edu.udo.cs.yale.operator.OperatorException;
import edu.udo.cs.yale.operator.IllegalInputException;
import edu.udo.cs.yale.operator.IODescription;
import edu.udo.cs.yale.operator.IOObject;
import edu.udo.cs.yale.operator.MissingIOObjectException;
import edu.udo.cs.yale.operator.Value;
import edu.udo.cs.yale.operator.learner.Model;
import edu.udo.cs.yale.operator.performance.PerformanceVector;
import edu.udo.cs.yale.operator.performance.PerformanceCriterion;

import java.util.List;

/** Abstract superclass of operator chains that split an {@link ExampleSet} into a training
 *  and test set and return a performance vector. The two inner operators must be a 
 *  learner returning a {@link Model} and an operator or operator chain that can
 *  apply this model and returns a {@link PerformanceVector}. Hence the second inner operator
 *  usually is an operator chain containing a model applier and a performance evaluator.
 *
 *  @author ingo, simon
 *  @version $Id: ValidationChain.java,v 1.5 2004/10/07 20:31:00 ingomierswa Exp $
 */
public abstract class ValidationChain extends OperatorChain {

    private static final Class[] OUTPUT_CLASSES = { AverageVector.class };
    private static final Class[] INPUT_CLASSES =  { ExampleSet.class };

    private PerformanceCriterion lastPerformance;
    private IOContainer learnResult;
    
    private boolean methodEvaluation;

    public ValidationChain() {
	addValue(new Value("performance", "The last performance average (main criterion).") {
		public double getValue() {
		    if (lastPerformance != null)
			return lastPerformance.getValue();
		    else
			return Double.NaN;
		}
	    });
	addValue(new Value("variance", "The variance of the last performance (main criterion).") {
		public double getValue() {
		    if (lastPerformance != null)
			return lastPerformance.getVariance();
		    else
			return Double.NaN;
		}
	    });
    }

    /** Returns the maximum number of innner operators. */
    public int getMaxNumberOfInnerOperators() { return 2; }
    /** Returns the minimum number of innner operators. */
    public int getMinNumberOfInnerOperators() { return 2; }

    /** returns the the classes this operator expects as input. */
    public Class[] getOutputClasses() { return OUTPUT_CLASSES; }
    /** returns the the classes this operator provides as output. */
    public Class[] getInputClasses() { return INPUT_CLASSES; }

    /** Checks the correctness of the input and output classes requested and provided,
     *  respectively, by the encapsulated inner operators of the <code>ValidationChain</code>.
     *  These input and output classes are OK, if the first inner operator returns a model and
     *  the second returns a performance vector.
     *  The method returns the output classes of the second encapsulated inner operator.
     */
    public Class[] checkIO(Class[] input) throws IllegalInputException {
	Operator learner   = getLearner();
	Operator evaluator = getEvaluator();
	input = learner.checkIO(input);
	if (!IODescription.containsClass(Model.class, input))
	    throw new IllegalInputException(this, learner, Model.class);

	Class[] newInput = new Class[input.length+1];
	for (int i = 0; i < input.length; i++) {
	    newInput[i] = input[i];
	}
	newInput[newInput.length-1] = ExampleSet.class;
	input = evaluator.checkIO(newInput);

	if (!IODescription.containsClass(AverageVector.class, input))
	    throw new IllegalInputException(this, evaluator, AverageVector.class);

	return input;
    }

    /** Returns the first encapsulated inner operator (or operator chain),
     *  i.e. the learning operator (chain). */
    private Operator getLearner() { return getOperator(0); }

    /** Returns the second encapsulated inner operator (or operator chain),
     *  i.e. the application and evaluation operator (chain) */
    private Operator getEvaluator() { return getOperator(1); }	

    /** Can be used by subclasses to set the performance of the example set. */
    protected void setResult(PerformanceCriterion pc) { lastPerformance = pc; }

    /** Applies the learner (= first encapsulated inner operator).
     *  This method recalculates the attribute statistics on the training set. */
    protected IOContainer learn(ExampleSet trainingSet) throws OperatorException {
	trainingSet.recalculateAllAttributeStatistics();
	return learnResult = getLearner().apply(getInput().prepend(new IOObject[] { trainingSet }));
    }

    /** Applies the applier and evaluator (= second encapsulated inner operator). 
     *  In order to reuse possibly created predicted label attributes, we do the following:
     *  We compare the predicted label of <code>testSet</code> before and
     *  after applying the inner operator. If it changed, the predicted label is removed again.
     *  No outer operator could ever see it. 
     *  This method recalculates the attribute statistics on the test set. */
    protected IOContainer evaluate(ExampleSet testSet) throws OperatorException {
	if (learnResult == null) {
	    throw new RuntimeException("Wrong use of ValidationChain.evaluate(ExampleSet): " +
				       "No preceding invocation of learn(ExampleSet)!");
	}
	testSet.recalculateAllAttributeStatistics();
  	Attribute predictedBefore = testSet.getPredictedLabel();
	IOContainer evalInput = learnResult.append(new IOObject[] { testSet });
	IOContainer result = getEvaluator().apply(evalInput);
  	Attribute predictedAfter = testSet.getPredictedLabel();
  	if ((predictedAfter != null) &&
  	    ((predictedBefore == null) ||
  	     (predictedBefore.getIndex() != predictedAfter.getIndex()))) {
  	    testSet.clearPredictedLabel();
  	    testSet.getExampleTable().removeAttribute(predictedAfter);	    
  	}
	learnResult = null;
	return result;
    }


    /** Searches for the average vectors in the given IOContainer and fills the list if it is empty 
     *  or build the averages.*/
    protected void handleAverages(IOContainer evalOutput, List averageVectors) {
	PerformanceVector performanceVector = null;
	int n = 0;
	boolean inputOk = true;
	while (inputOk) {
	    try {
		AverageVector currentAverage = (AverageVector)evalOutput.getInput(AverageVector.class);
		if (averageVectors.size() == 0) { 
                    // first run --> do not calculate average values but fill the vector list
		    averageVectors.add(currentAverage);
		} else {  
                    // later runs --> build the average with corresponding average vectors
		    AverageVector oldVector = (AverageVector)averageVectors.get(n++);
		    for (int i = 0; i  < oldVector.size(); i++) {
			oldVector.getAveragable(i).buildAverage(currentAverage.getAveragable(i));
		    }
		}
	    } catch (MissingIOObjectException e) {
		inputOk = false;
	    }
	}
    }

    /** Returns the first performance vector in the given list or null if no performance vectors exist. */
    protected PerformanceVector getPerformanceVector(List averageVectors) {
	java.util.Iterator i = averageVectors.iterator();
	while (i.hasNext()) {
	    AverageVector currentAverage = (AverageVector)i.next();
	    if (currentAverage instanceof PerformanceVector)
		return (PerformanceVector)currentAverage;
	}
	return null;
    }

    public abstract int getNumberOfValidationSteps();

    public int getNumberOfSteps() {
	return getNumberOfValidationSteps() * super.getNumberOfChildrensSteps() + 1;
    }
}