budgetedlearningcurveresultproducer.java

wekaUT是 university texas austin 开发的基于weka的半指导学习(semi supervised learning)的分类器

/*
 *    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.
 */

/*
 *    BudgetedLearningCurveResultProducer.java
 *    Copyright (C) 2005 Prem Melville
 *
 */

////////////////////////////////
//
// WARNING: UNDER DEVELOPMENT
//
////////////////////////////////

package weka.experiment;

import java.util.*;
import java.io.*;
import weka.classifiers.*;
import weka.core.*;


/**
 * Does an N-fold cross-validation, but generates a learning curve by
 * also varying the number of training examples. Creates a split that
 * uses increasingly larger fractions of the full training set from
 * the N fold but always using the same N-fold test set for
 * testing. If this is applied to an active learner, then the training
 * examples are selected actively by the learner from the pool of
 * unlabeled examples. If this is not used with an active learner, it
 * should produce the same results as the
 * LearningCurveCrossValidationResultProducer.
 *
 * @author Prem Melville ([e-mail]) 
 */

public class BudgetedLearningCurveResultProducer 
    implements ResultProducer, OptionHandler, AdditionalMeasureProducer {

    /** 
     * The directory used when loading feature cost files on demand, null indicates
     * current directory 
     */
    protected File m_FeatureCostDirectory = new File(System.getProperty("user.dir")+"/featureCosts");
    
    /** Costs for acquiring each feature */ 
    protected double []m_FeatureCosts = null;

    /** The dataset of interest */
    protected Instances m_Instances;

    /** The ResultListener to send results to */
    protected ResultListener m_ResultListener = new CSVResultListener();

    /** The number of folds in the cross-validation */
    protected int m_NumFolds = 10;

    /** Save raw output of split evaluators --- for debugging purposes */
    protected boolean m_debugOutput = false;

    /** The output zipper to use for saving raw splitEvaluator output */
    protected OutputZipper m_ZipDest = null;

    /** The destination output file/directory for raw output */
    protected File m_OutputFile = new File(
					   new File(System.getProperty("user.dir")), 
					   "splitEvalutorOut.zip");

    /** The SplitEvaluator used to generate results */
    //protected SplitEvaluator m_SplitEvaluator = new ClassifierSplitEvaluator();
    protected SplitEvaluator m_SplitEvaluator = new FeatureCostSensitiveClassifierSplitEvaluator();

    /** The names of any additional measures to look for in SplitEvaluators */
    protected String [] m_AdditionalMeasures = null;

    /** 
     * The minimum number of instances to use. If this is zero, the first
     * step will contain m_StepSize instances 
     */
    protected double m_LowerSize = 0;
  
    /**
     * The maximum number of instances to use. -1 indicates no maximum 
     * (other than the total number of instances)
     */
    protected double m_UpperSize = -1;

    /** The number of instances to add at each step */
    protected double m_StepSize = 1;

    /** The specific points to plot, either integers representing specific numbers of training examples,
     * or decimal fractions representing percentages of the full training set*/
    protected double[] m_PlotPoints;

    /** The current dataset size during stepping */
    protected int m_CurrentSize = 0;

    /** Cost to build current model */
    protected double m_Cost = 0.0;
    
    /* The name of the key field containing the dataset name */
    public static String DATASET_FIELD_NAME = "Dataset";

    /* The name of the key field containing the run number */
    public static String RUN_FIELD_NAME = "Run";

    /* The name of the key field containing the fold number */
    public static String FOLD_FIELD_NAME = "Fold";

    /* The name of the result field containing the timestamp */
    public static String TIMESTAMP_FIELD_NAME = "Date_time";
    
    /* The name of the result field containing the sample selection time */
    public static String SELECTION_TIME_FIELD_NAME = "Selection_time";
    
    /* The name of the key field containing the learning rate step number */
    public static String STEP_FIELD_NAME = "Total_instances";

    /* The name of the key field containing the fraction of total instances used */
    public static String FRACTION_FIELD_NAME = "Fraction_instances";
    
    public static String COST_FIELD_NAME = "Model_cost";
    
    /* Extension for feature cost files */
    public static String FEATURE_COST_FILE_EXTENSION = ".featureCosts";
    
    /* Indicates whether fractions or actual number of instances have been specified */
    protected boolean m_IsFraction = false;
    
    /**
     * Returns a string describing this result producer
     * @return a description of the result producer suitable for
     * displaying in the explorer/experimenter gui
     */
    public String globalInfo() {
	return "Performs a learning-curve cross validation run using a supplied "
	    +"split evaluator. Trains on increasing subsets of the training data for each split, "
	    +"repeatedly testing on the test set for that split after training on subsets of various sizes.";
    }
    
    

    /**
     * Sets the dataset that results will be obtained for.
     *
     * @param instances a value of type 'Instances'.
     */
    public void setInstances(Instances instances) {
    
	m_Instances = instances;
	loadFeatureCosts();
    }
    
    
    
    /**
     * Returns the directory that will be searched for feature cost files when
     * loading on demand.
     *
     * @return The cost file search directory.
     */
    public File getFeatureCostDirectory() {
	return m_FeatureCostDirectory;
    }
    
    /**
     * Sets the directory that will be searched for feature cost files when
     * loading on demand.
     *
     * @param newDir The cost file search directory.
     */
    public void setFeatureCostDirectory(File newDir) {
	if (newDir.isDirectory()) {
	    m_FeatureCostDirectory = newDir;
	} else {
	    m_FeatureCostDirectory = new File(newDir.getParent());
	}
    }
    
    //Read feature costs from file
    protected void loadFeatureCosts(){
	try{
	    String costName = m_Instances.relationName() + FEATURE_COST_FILE_EXTENSION;
	    File costFile = new File(getFeatureCostDirectory(), costName);
	    if (!costFile.exists()) {
		throw new Exception("Feature cost file doesn't exist: " + costFile);
	    }
	    int numFeatures = numFeatures();
	    m_FeatureCosts = new double[numFeatures];
	    
	    BufferedReader bin = new BufferedReader(new FileReader(costFile));
	    String line = bin.readLine();//read first line - ignore rest of file
	    bin.close();
	    StringTokenizer st = new StringTokenizer(line);
	    int ctr = 0;//count number of costs in the file
	    while(st.hasMoreTokens()) {
		if(ctr < numFeatures){
		    m_FeatureCosts[ctr] = Double.parseDouble(st.nextToken());
		    ctr++;
		}else{
		    //throw new Exception("Feature cost file has too many costs.");   
		    System.err.println("Feature cost file has too many costs: "+costFile);   
		    System.exit(0);
		}
	    }
	    if(ctr < numFeatures){
		System.err.println("Feature cost file has too few costs: "+costFile);   
		System.exit(0);
		//throw new Exception("Feature cost file has too few costs.");   
	    }
	}catch (Exception e){
	    System.err.println (e);
	}
	//DEBUG
	Utils.printArray(m_FeatureCosts);
    }
    
    /**
     * Sets the object to send results of each run to.
     *
     * @param listener a value of type 'ResultListener'
     */
    public void setResultListener(ResultListener listener) {

	m_ResultListener = listener;
    }

    
    /**
     * Set a list of method names for additional measures to look for
     * in SplitEvaluators. This could contain many measures (of which only a
     * subset may be produceable by the current SplitEvaluator) if an experiment
     * is the type that iterates over a set of properties.
     * @param additionalMeasures an array of measure names, null if none
     */
    public void setAdditionalMeasures(String [] additionalMeasures) {
	m_AdditionalMeasures = additionalMeasures;

	if (m_SplitEvaluator != null) {
	    System.err.println("LearningCurveCrossValidationResultProducer: setting additional "
			       +"measures for "
			       +"split evaluator");
	    m_SplitEvaluator.setAdditionalMeasures(m_AdditionalMeasures);
	}
    }

    /**
     * Returns an enumeration of any additional measure names that might be
     * in the SplitEvaluator
     * @return an enumeration of the measure names
     */
    public Enumeration enumerateMeasures() {
	Vector newVector = new Vector();
	if (m_SplitEvaluator instanceof AdditionalMeasureProducer) {
	    Enumeration en = ((AdditionalMeasureProducer)m_SplitEvaluator).
		enumerateMeasures();
	    while (en.hasMoreElements()) {
		String mname = (String)en.nextElement();
		newVector.addElement(mname);
	    }
	}
	return newVector.elements();
    }
  
    /**
     * Returns the value of the named measure
     * @param measureName the name of the measure to query for its value
     * @return the value of the named measure
     * @exception IllegalArgumentException if the named measure is not supported
     */
    public double getMeasure(String additionalMeasureName) {
	if (m_SplitEvaluator instanceof AdditionalMeasureProducer) {
	    return ((AdditionalMeasureProducer)m_SplitEvaluator).
		getMeasure(additionalMeasureName);
	} else {
	    throw new IllegalArgumentException("LearningCurveCrossValidationResultProducer: "
					       +"Can't return value for : "+additionalMeasureName
					       +". "+m_SplitEvaluator.getClass().getName()+" "
					       +"is not an AdditionalMeasureProducer");
	}
    }
  
    /**
     * Gets a Double representing the current date and time.
     * eg: 1:46pm on 20/5/1999 -> 19990520.1346
     *
     * @return a value of type Double
     */
    public static Double getTimestamp() {

	Calendar now = Calendar.getInstance(TimeZone.getTimeZone("UTC"));
	double timestamp = now.get(Calendar.YEAR) * 10000
	    + (now.get(Calendar.MONTH) + 1) * 100
	    + now.get(Calendar.DAY_OF_MONTH)
	    + now.get(Calendar.HOUR_OF_DAY) / 100.0
	    + now.get(Calendar.MINUTE) / 10000.0;
	return new Double(timestamp);
    }
  
    /**
     * Prepare to generate results.
     *
     * @exception Exception if an error occurs during preprocessing.
     */
    public void preProcess() throws Exception {

	if (m_SplitEvaluator == null) {
	    throw new Exception("No SplitEvalutor set");
	}
	if (m_ResultListener == null) {
	    throw new Exception("No ResultListener set");
	}
	m_ResultListener.preProcess(this);
    }
  
    /**
     * Perform any postprocessing. When this method is called, it indicates
     * that no more requests to generate results for the current experiment
     * will be sent.
     *
     * @exception Exception if an error occurs
     */
    public void postProcess() throws Exception {

	m_ResultListener.postProcess(this);

	if (m_debugOutput) {
	    if (m_ZipDest != null) {
		m_ZipDest.finished();
		m_ZipDest = null;
	    }
	}
    }
  
    /**
     * Gets the keys for a specified run number. Different run
     * numbers correspond to different randomizations of the data. Keys
     * produced should be sent to the current ResultListener
     *
     * @param run the run number to get keys for.
     * @exception Exception if a problem occurs while getting the keys
     */
    public void doRunKeys(int run) throws Exception {
	int numExtraKeys;
	if(m_IsFraction)
	    numExtraKeys = 5;
	else numExtraKeys = 4;
	
	if (m_Instances == null) {
	    throw new Exception("No Instances set");
	}
	if (m_ResultListener == null) {
	    throw new Exception("No ResultListener set");
	}
	for (int fold = 0; fold < m_NumFolds; fold++) {
	    int pointNum = 0;
	    // For each subsample size
	    if (m_PlotPoints != null) {
		m_CurrentSize = plotPoint(0);
	    }
	    else if (m_LowerSize == 0) {
		m_CurrentSize = stepSize(m_StepSize);
	    } else {
		m_CurrentSize = lowerSize(m_LowerSize);
	    }
	    while (m_CurrentSize <= maxNumQueries()) {
		// Add in some fields to the key like run and fold number, dataset name
		Object [] seKey = m_SplitEvaluator.getKey();
		Object [] key = new Object [seKey.length + numExtraKeys];
		key[0] = Utils.backQuoteChars(m_Instances.relationName());
		key[1] = "" + run;
		key[2] = "" + (fold + 1);
		key[3] = "" + m_CurrentSize;
		if(m_IsFraction) key[4] = "" + m_PlotPoints[pointNum];
		System.arraycopy(seKey, 0, key, numExtraKeys, seKey.length);
		if (m_ResultListener.isResultRequired(this, key)) {
		    try {
			m_ResultListener.acceptResult(this, key, null);
		    } catch (Exception ex) {
			// Save the train and test datasets for debugging purposes?
			throw ex;
		    }
		}
		if (m_PlotPoints != null) {
		    pointNum ++;
		    m_CurrentSize = plotPoint(pointNum);
		}
		else {
		    m_CurrentSize += stepSize(m_StepSize);
		}
	    }
	}
    }

    
    /** 
     * Get the maximum number of queries base on the upperSize limit
     * or maximum training set size from the n-fold CV 
     */
    protected int maxNumQueries() {
	if (m_UpperSize == -1 || m_PlotPoints != null)
	    return ((int)(m_Instances.numInstances()*(1 - 1/((double)m_NumFolds))))*numFeatures();
	else if(isInteger(m_UpperSize))
	    return (int) m_UpperSize;
	else return (int) (m_UpperSize*((int)(m_Instances.numInstances()*(1 - 1/((double)m_NumFolds))))*numFeatures());
    }
    
    //If lowersize is a fraction, compute as a fraction of total number of features.
    protected int lowerSize(double lowerSize){
	if(isInteger(lowerSize))
	    return (int) lowerSize;
	else return (int) (lowerSize * ((int)(m_Instances.numInstances()*(1 - 1/((double)m_NumFolds))))*numFeatures());
    }
    
    //If lowersize is a fraction, compute as a fraction of total number of features.
    protected int stepSize(double stepSize){
	int out;
	if(isInteger(stepSize))
	    out = (int) stepSize;
	else out = (int) (stepSize * ((int)(m_Instances.numInstances()*(1 - 1/((double)m_NumFolds))))*numFeatures());
	//step size must at least be 1
	if(out<1) out=1;
	return out;
    }
    
    //Returns the number of features in the datasets (excluding the class attribute)
    protected int numFeatures(){
	return (m_Instances.numAttributes() - 1);
    }
    
    /**
     * Gets the results for a specified run number. Different run
     * numbers correspond to different randomizations of the data. Results
     * produced should be sent to the current ResultListener
     *
     * @param run the run number to get results for.
     * @exception Exception if a problem occurs while getting the results
     */