segment.java

CRF1.2

package iitb.Segment;
import java.io.*;
import java.util.*;
import iitb.CRF.*;
import iitb.Model.*;
import iitb.Utils.*;
/**
 *
 * @author Sunita Sarawagi
 *
 */ 



public class Segment {
    String inName;
    String outDir;
    String baseDir="";
    int nlabels;

    String delimit=" \t"; // used to define token boundaries
    String tagDelimit="|"; // seperator between tokens and tag number
    String impDelimit=""; // delimiters to be retained for tagging
    String groupDelimit=null;

    boolean confuseSet[]=null;
    boolean validate = false; 
    String mapTagString = null;
    String smoothType = "";
        
    String modelArgs = "";
    String featureArgs = "";
    String modelGraphType = "naive";

    LabelMap labelMap;
    Options options;

    CRF crfModel;
    FeatureGenImpl featureGen;
    public FeatureGenerator featureGenerator() {return featureGen;}

    public static void main(String argv[]) throws Exception {
	if (argv.length < 3) {
	    System.out.println("Usage: java Tagger train|test|calc -f <conf-file>");
	    return;
	}
	Segment segment = new Segment();
	segment.parseConf(argv);
	if (argv[0].toLowerCase().equals("all")) {
	    segment.train();
	    segment.doTest();
	    segment.calc();
	} 
	if (argv[0].toLowerCase().equals("train")) {
	    segment.train();
	} 
	if (argv[0].toLowerCase().equals("test")) {
	    segment.test();
	}
	if (argv[0].toLowerCase().equals("calc")) {
	    segment.calc();
	}
    }

    public void parseConf(String argv[]) throws Exception {
	options = new Options();
	int startIndex = 1;
	if ((argv.length >= 2) && argv[1].toLowerCase().equals("-f")) {
	    options.load(new FileInputStream(argv[2]));
	}
	options.add(3, argv);
	processArgs();
    }

    public void processArgs() throws Exception {
	String value = null;
	if ((value = options.getMandatoryProperty("numlabels")) != null) {
	    nlabels=Integer.parseInt(value);
	}
	if ((value = options.getProperty("binary")) != null) {
	    nlabels = 2;
	    labelMap = new BinaryLabelMap(options.getInt("binary"));
	} else {
	    labelMap = new LabelMap();
	}
	if ((value = options.getMandatoryProperty("inname")) != null) {
	    inName=new String(value);
	}
	if ((value = options.getMandatoryProperty("outdir")) != null) {
	    outDir=new String(value);
	}
	if ((value = options.getProperty("basedir")) != null) {
	    baseDir=new String(value);
	}
	if ((value = options.getProperty("tagdelimit")) != null) {
	    tagDelimit=new String(value);
	}
	// delimiters that will be ignored.
	if ((value = options.getProperty("delimit")) != null) {
	    delimit=new String(value);
	}
	if ((value = options.getProperty("impdelimit")) != null) {
	    impDelimit=new String(value);
	}
	if ((value = options.getProperty("groupdelimit")) != null) {
	    groupDelimit=value;
	}
	if ((value = options.getProperty("confusion")) != null) {
	    StringTokenizer confuse=new StringTokenizer(value,", ");
	    int confuseSize=confuse.countTokens();
	    confuseSet=new boolean[nlabels+1];
	    for(int i=0 ; i<confuseSize ; i++) {
		confuseSet[Integer.parseInt(confuse.nextToken())]=true;
	    }
	}
	if ((value = options.getProperty("map-tags")) != null) {
	    mapTagString = value;
	}
	if ((value = options.getProperty("validate")) != null) {
	    validate = true;
	}
	if ((value = options.getProperty("model-args")) != null) {
	    modelArgs = value;
	    System.out.println(modelArgs);
	}
	if ((value = options.getProperty("feature-args")) != null) {
	    featureArgs = value;
	}
	if ((value = options.getProperty("modelGraph")) != null) {
	    modelGraphType = value;
	}
    }
    void  allocModel() throws Exception {
	// add any code related to dependency/consistency amongst paramter
	// values here..
	if (modelGraphType.equals("semi-markov")) {
	    if (options.getInt("debugLvl") > 1) {
		Util.printDbg("Creating semi-markov model");
	    }
	    NestedFeatureGenImpl nfgen = new NestedFeatureGenImpl(nlabels,options);
	    featureGen = nfgen;
	    crfModel = new NestedCRF(featureGen.numStates(),nfgen,options);
	} else 
	{
	    featureGen = new FeatureGenImpl(modelGraphType, nlabels);
	    crfModel=new CRF(featureGen.numStates(),featureGen,options);
        }
    }
    class TestRecord implements SegmentDataSequence {
    	String seq[];
    	int path[];
    	TestRecord(String s[]) {
    	    seq=s;
    	    path=new int[seq.length];
    	}
    	void init(String s[]) {
    	    seq = s;
    	    if ((path == null) || (path.length < seq.length)) {
    		path = new int[seq.length];
    	    }
    	}
    	public void set_y(int i, int l) {path[i] = l;} // not applicable for training data.
    	public int y(int i) {return path[i];}
    	public int length() {return seq.length;}
    	public Object x(int i) {return seq[i];}
    	/* (non-Javadoc)
    	 * @see iitb.CRF.SegmentDataSequence#getSegmentEnd(int)
    	 */
    	public int getSegmentEnd(int segmentStart) {
    		if ((segmentStart > 0) && (y(segmentStart) == y(segmentStart-1)))
    			return -1;
    		for (int i = segmentStart+1; i < length(); i++) {
    			if (y(i)!= y(segmentStart))
    				return i-1;
    		}
    		return length()-1;
    	}
    	/* (non-Javadoc)
    	 * @see iitb.CRF.SegmentDataSequence#setSegment(int, int, int)
    	 */
    	public void setSegment(int segmentStart, int segmentEnd, int y) {
    		for (int i = segmentStart; i <= segmentEnd; i++)
    			set_y(i,y);
    	}
        };

 
    public int[] segment(TestRecord testRecord, int[] groupedToks, String collect[]) {
	for (int i = 0; i < testRecord.length(); i++)
	    testRecord.seq[i] = AlphaNumericPreprocessor.preprocess(testRecord.seq[i]);
	crfModel.apply(testRecord);
	featureGen.mapStatesToLabels(testRecord);
	int path[] = testRecord.path;
	for (int i = 0; i < nlabels; i++)
	    collect[i] = null;
	for(int i=0 ; i<testRecord.length() ; i++) {
	    // System.out.println(testRecord.seq[i] + " " + path[i]);
	    int snew=path[i];
	    if (snew >= 0) {
		if (collect[snew]==null) {
		    collect[snew]=testRecord.seq[i];
		} else {
		    collect[snew]=collect[snew]+" "+testRecord.seq[i];
		}
	    }
	}
	return path;
    }
    public void train() throws Exception {
	DataCruncher.createRaw(baseDir+"/data/"+inName+"/"+inName+".train",tagDelimit);
	File dir=new File(baseDir+"/learntModels/"+outDir);
	dir.mkdirs();
	TrainData trainData = DataCruncher.readTagged(nlabels,baseDir+"/data/"+inName+"/"+inName+".train",baseDir+"/data/"+inName+"/"+inName+".train",delimit,tagDelimit,impDelimit,labelMap);
	AlphaNumericPreprocessor.preprocess(trainData,nlabels);

	allocModel();
	featureGen.train(trainData);
	double featureWts[] = crfModel.train(trainData);
	if (options.getInt("debugLvl") > 1) {
	    Util.printDbg("Training done");
	}
	crfModel.write(baseDir+"/learntModels/"+outDir+"/crf");
	featureGen.write(baseDir+"/learntModels/"+outDir+"/features");
	if (options.getInt("debugLvl") > 1) {
	    Util.printDbg("Writing model to "+ baseDir+"/learntModels/"+outDir+"/crf");
	}
	if (options.getProperty("showModel") != null) {
	    featureGen.displayModel(featureWts);
	}
    }

    public void test() throws Exception {
	allocModel();
	featureGen.read(baseDir+"/learntModels/"+outDir+"/features");
	crfModel.read(baseDir+"/learntModels/"+outDir+"/crf");
	doTest();
    }
    public void doTest() throws Exception {
	File dir=new File(baseDir+"/out/"+outDir);
	dir.mkdirs();
	TestData testData = new TestData(baseDir+"/data/"+inName+"/"+inName+".test",delimit,impDelimit,groupDelimit);
	TestDataWrite tdw = new TestDataWrite(baseDir+"/out/"+outDir+"/"+inName+".test",baseDir+"/data/"+inName+"/"+inName+".test",delimit,tagDelimit,impDelimit,labelMap);
	
	String collect[] = new String[nlabels];
	TestRecord testRecord = new TestRecord(collect);
	for(String seq[] = testData.nextRecord();  seq != null; 
	    seq = testData.nextRecord()) {
	    testRecord.init(seq);
	    if (options.getInt("debugLvl") > 1) {
		Util.printDbg("Invoking segment on " + seq);
	    }
	    int path[] = segment(testRecord, testData.groupedTokens(), collect);
	    tdw.writeRecord(path,testRecord.length());
	}
	tdw.close();
    }
    TrainData taggedData = null;
    int[] allLabels(TrainRecord tr) {
	int[] labs = new int[tr.length()];
	for (int i = 0; i < labs.length; i++)
	    labs[i] = tr.y(i);
	return labs;
    }
    String arrayToString(Object[] ar) {
	String st = "";
	for (int i = 0; i < ar.length; i++)
	    st += (ar[i] + " ");
	return st;
    }
    public void calc() throws Exception {
	Vector s=new Vector();
	TrainData tdMan = DataCruncher.readTagged(nlabels,baseDir+"/data/"+inName+"/"+inName+".test",baseDir+"/data/"+inName+"/"+inName+".test",delimit,tagDelimit,impDelimit,labelMap);
	TrainData tdAuto = DataCruncher.readTagged(nlabels,baseDir+"/out/"+outDir+"/"+inName+".test",baseDir+"/data/"+inName+"/"+inName+".test",delimit,tagDelimit,impDelimit,labelMap);
	DataCruncher.readRaw(s,baseDir+"/data/"+inName+"/"+inName+".test","","");
	int len=tdAuto.size();
	int truePos[]=new int[nlabels+1];
	int totalMarkedPos[]=new int[nlabels+1];
	int totalPos[]=new int[nlabels+1];
	int confuseMatrix[][]=new int[nlabels][nlabels];
	boolean printDetails = (options.getInt("debugLvl") > 0);
	if (tdAuto.size() != tdMan.size()) {
	    // Sanity Check
	    System.out.println("Length Mismatch - Raw: "+len+" Auto: "+tdAuto.size()+" Man: "+tdMan.size());
	}
		
	for(int i=0 ; i<len ; i++) {
	    String raw[]=(String [])(s.get(i));
	    TrainRecord trMan = tdMan.nextRecord();
	    TrainRecord trAuto = tdAuto.nextRecord();
	    int tokenMan[]=allLabels(trMan);
	    int tokenAuto[]=allLabels(trAuto);

	    if (tokenMan.length!=tokenAuto.length) {
		// Sanity Check
		System.out.println("Length Mismatch - Manual: "+tokenMan.length+" Auto: "+tokenAuto.length);
		//			continue;
	    }
	    // remove invalid tagging.
	    boolean invalidMatch = false;
	    int tlen=tokenMan.length;
	    for (int j = 0; j < tlen; j++) {
		if (printDetails) System.err.println(tokenMan[j] + " " + tokenAuto[j]);
		if (tokenAuto[j] < 0) {
		    invalidMatch = true;
		    break;
		}
	    }
	    if (invalidMatch) {
		if (printDetails) System.err.println("No valid path");
		continue;
	    }
	    int correctTokens=0;
	    for(int j=0 ; j<tlen ; j++) {
		totalMarkedPos[tokenAuto[j]]++;
		totalMarkedPos[nlabels]++;
		totalPos[tokenMan[j]]++;
		totalPos[nlabels]++;
		confuseMatrix[tokenMan[j]][tokenAuto[j]]++;
		if (tokenAuto[j]==tokenMan[j]) {
		    correctTokens++;
		    truePos[tokenMan[j]]++;
		    truePos[nlabels]++;
		}
	    }
	    if (printDetails) System.err.println("Stats: "+correctTokens+" "+(tlen));
	    int rlen=raw.length;
	    for(int j=0 ; j<rlen ; j++) {
		if (printDetails) System.err.print(raw[j]+" ");
	    }
	    if (printDetails) System.err.println();
	    for(int j=0 ; j<nlabels ; j++) {			    
		String mstr = "";
		for (int k = 0;  k < trMan.numSegments(j);k++)
		    mstr += arrayToString(trMan.tokens(j,k));
		String astr = "";
		for (int k = 0;  k < trAuto.numSegments(j);k++)
		    astr += arrayToString(trAuto.tokens(j,k));

		if (! mstr.equalsIgnoreCase(astr))
		    if (printDetails) System.err.print("W");
		if (printDetails) System.err.println(j+": "+ mstr+" : "+astr);
	    }
	    if (printDetails) System.err.println();
	}

	if (confuseSet!=null) {
	    System.out.println("Confusion Matrix:");
	    System.out.print("M\\A");
	    for(int i=0 ; i<nlabels ; i++) {
		if (confuseSet[i]) {
		    System.out.print("\t"+(i));
		}
	    }
	    System.out.println();
	    for(int i=0 ; i<nlabels ; i++) {
		if (confuseSet[i]) {
		    System.out.print(i);
		    for(int j=0 ; j<nlabels ; j++) {
			if (confuseSet[j]) {
			    System.out.print("\t"+confuseMatrix[i][j]);
			}
		    }
		    System.out.println();
		}
	    }
	}
	System.out.println("\n\nCalculations:");
	System.out.println();
	System.out.println("Label\tTrue+\tMarked+\tActual+\tPrec.\tRecall\tF1");
	double prec,recall;
	for(int i=0 ; i<nlabels ; i++) {
	    prec=(totalMarkedPos[i]==0)?0:((double)(truePos[i]*100000/totalMarkedPos[i]))/1000;
	    recall=(totalPos[i]==0)?0:((double)(truePos[i]*100000/totalPos[i]))/1000;
	    System.out.println((i)+":\t"+truePos[i]+"\t"+totalMarkedPos[i]+"\t"+totalPos[i]+"\t"+prec+"\t"+recall+"\t"+2*prec*recall/(prec+recall));
	}
	System.out.println("---------------------------------------------------------");
	prec=(totalMarkedPos[nlabels]==0)?0:((double)(truePos[nlabels]*100000/totalMarkedPos[nlabels]))/1000;
	recall=(totalPos[nlabels]==0)?0:((double)(truePos[nlabels]*100000/totalPos[nlabels]))/1000;
	System.out.println("Ov:\t"+truePos[nlabels]+"\t"+totalMarkedPos[nlabels]+"\t"+totalPos[nlabels]+"\t"+prec+"\t"+recall+"\t"+2*prec*recall/(prec+recall));

    }
};