expsinglelm.java

用于multivariate时间序列分类

/**
 *  Single classifier solution. That is to say, we cluster all the instances
 *  using the same clustering algorithms. 
 * 
 * This is a "low memory" version, written to see if memory consumption can be reduced. 
 * 
 * @author Waleed Kadous
 * @version $Id: ExpSingleLM.java,v 1.1.1.1 2002/06/28 07:36:16 waleed Exp $
 */
package tclass;  

import tclass.clusteralg.*; 
import tclass.util.*; 
// import tclass.learnalg.*; 
import weka.classifiers.*; 
import weka.classifiers.j48.*; 
import weka.attributeSelection.*; 
import weka.filters.*; 
import weka.core.*; 
import java.io.*; 
import java.util.*; 


public class ExpSingleLM {
    // Ok. What we are going to do is to separate the learning task in 
    // an interesting way. 
    // First of all, though, the standard stuff
    
    static Runtime rt = Runtime.getRuntime(); 
    String domDescFile = "sl.tdd"; 
    String trainDataFile = "sl.tsl"; 
    String testDataFile = "sl.ttl"; 
    // String globalDesc = "test._gc"; 
    // String evExtractDesc = "test._ee";
    String evClusterDesc = "test._ec"; 
    String settingsFile = "test.tal"; 
    String learnerStuff = "weka.classifiers.j48.J48"; 
    boolean featureSel = false; 
    boolean makeDesc = false; 
    boolean trainResults = false; 
    void parseArgs(String[] args){
        for(int i=0; i < args.length; i++){
            if(args[i].equals("-tr")){
                trainDataFile = args[++i]; 
            }
            if(args[i].equals("-dd")){
                domDescFile = args[++i]; 
            }
            if(args[i].equals("-te")){
                testDataFile = args[++i];
            }
            if(args[i].equals("-settings")){
                settingsFile = args[++i]; 
            }
            if(args[i].equals("-fs")){
                featureSel = true; 
            }
            if(args[i].equals("-md")){
                makeDesc = true; 
            }
            if(args[i].equals("-trainres")){
                trainResults = true; 
            }
            if(args[i].equals("-l")){
                learnerStuff = args[++i]; 
                learnerStuff = learnerStuff.replace(':', ' '); 
                System.err.println("Learner String is: " + learnerStuff); 
            }
        }
    }

    // Alright. This is downright funky hacky stuff. 

    public static void mem(String label){
        System.out.println("Memory at checkpt " + label + ": " + (rt.totalMemory()/1024/1024) + " megabytes."); 
    }

    public static void main(String[] args) throws Exception {
        Debug.setDebugLevel(Debug.PROGRESS); 
        ExpSingleLM thisExp = new ExpSingleLM(); 
        thisExp.parseArgs(args); 
        mem("PARSE"); 
        DomDesc domDesc = new DomDesc(thisExp.domDescFile); 
        ClassStreamVecI trainStreamData = new
            ClassStreamVec(thisExp.trainDataFile, domDesc); 
        Debug.dp(Debug.PROGRESS, "PROGRESS: Training data read in");  
        mem("TRAINDATAIN"); 
        Settings settings = new Settings(thisExp.settingsFile, domDesc); 
        

        EventExtractor evExtractor = settings.getEventExtractor(); 
        // Global data is likely to be included in every model; so we
        // might as well calculated now
        GlobalCalc globalCalc = settings.getGlobalCalc(); 

        ClassStreamAttValVecI trainGlobalData =
            globalCalc.applyGlobals(trainStreamData);
        // And we might as well extract the events. 

        Debug.dp(Debug.PROGRESS, "PROGRESS: Training data globals calculated.");  
        mem("TRAINGLOBAL"); 
        Debug.dp(Debug.PROGRESS, "Train: " + trainGlobalData.size());  


        ClassStreamEventsVecI trainEventData =
            evExtractor.extractEvents(trainStreamData); 
        Debug.dp(Debug.PROGRESS, "PROGRESS: Training events extracted");  
        mem("EVENTEXTRACT"); 
        // System.out.println(trainEventData.toString()); 


        // Now we want the clustering algorithms only to cluster
        // instances of each class. Make an array of clusterers, 
        // one per class. 

        int numClasses = domDesc.getClassDescVec().size(); 
        EventDescVecI eventDescVec = evExtractor.getDescription(); 
        EventClusterer eventClusterer = settings.getEventClusterer(); 
        Debug.dp(Debug.PROGRESS, "PROGRESS: Data rearranged.");  
        mem("REARRANGE"); 

        //And now load it up. 
        StreamEventsVecI trainEventSEV =
            trainEventData.getStreamEventsVec(); 
        ClassificationVecI trainEventCV = trainEventData.getClassVec();
        int numTrainStreams = trainEventCV.size(); 
        ClusterVecI clusters = eventClusterer.clusterEvents(trainEventData); 
        Debug.dp(Debug.PROGRESS, "PROGRESS: Clustering complete"); 
        Debug.dp(Debug.PROGRESS, "Clusters are:"); 
        Debug.dp(Debug.PROGRESS, "\n" + eventClusterer.getMapping()); 
        Debug.dp(Debug.PROGRESS, "PROGRESS: Clustering complete. ");  
        mem("CLUSTER"); 

        // But wait! There's more! There is always more. 
        // The first thing was only useful for clustering. 
        // Now attribution. We want to attribute all the data. So we are going 
        // to have one dataset for each learner. 
        // First set up the attributors. 

        Attributor attribs = new Attributor(domDesc, clusters, 
                                            eventClusterer.getDescription()); 
        Debug.dp(Debug.PROGRESS, "PROGRESS: AttributorMkr complete."); 
        mem("MAKEATTRIBUTOR"); 


        ClassStreamAttValVecI trainEventAtts =attribs.attribute(trainStreamData, trainEventData); 
        Debug.dp(Debug.PROGRESS, "PROGRESS: Training data Attribution complete."); 
        mem("TRAINATTRIBUTION");

        // Combine all data sources. For now, globals go in every
        // one. 

        Combiner c = new Combiner(); 
        ClassStreamAttValVecI trainAtts = c.combine(trainGlobalData,
                                            trainEventAtts); 

        mem("TRAINCOMBINATION"); 
        trainStreamData = null; 
        trainEventSEV = null; 
        trainEventCV = null; 
        System.gc(); 
        mem("TRAINGC");

        // So now we have the raw data in the correct form for each
        // attributor. 
        // And now, we can construct a learner for each case. 
        // Well, for now, I'm going to do something completely crazy. 
        // Let's run each classifier nonetheless over the whole data
        // ... and see what the hell happens. Maybe some voting scheme 
        // is possible!! This is a strange form of ensemble
        // classifier. 
        // Each naive bayes algorithm only gets one 

        Debug.setDebugLevel(Debug.PROGRESS); 
        int[] selectedIndices = null;  
        String [] classifierSpec = Utils.splitOptions(thisExp.learnerStuff);
        if (classifierSpec.length == 0) {
            throw new Exception("Invalid classifier specification string");
        }        
        String classifierName = classifierSpec[0];
        classifierSpec[0] = "";
        Classifier learner = Classifier.forName(classifierName, classifierSpec);
        Debug.dp(Debug.PROGRESS, "PROGRESS: Beginning format conversion for class "); 
        Instances  data = WekaBridge.makeInstances(trainAtts, "Train ");
        Debug.dp(Debug.PROGRESS, "PROGRESS: Conversion complete. Starting learning");    
        mem("ATTCONVERSION");
        if(thisExp.featureSel){
                Debug.dp(Debug.PROGRESS, "PROGRESS: Doing feature selection");    
                BestFirst bfs = new BestFirst();
                CfsSubsetEval cfs = new CfsSubsetEval(); 
                cfs.buildEvaluator(data); 
                selectedIndices = bfs.search(cfs, data); 
                // Now extract the features. 
                System.err.print("Selected features: ");
                String featureString = new String(); 
                for(int j=0; j < selectedIndices.length; j++){
                    featureString += (selectedIndices[j] +1)+ ",";
                }
                featureString += ("last"); 
                System.err.println(featureString); 
               // Now apply the filter.