Index: trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaBaseTraining2.java
===================================================================
--- trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaBaseTraining2.java (revision 19)
+++ trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaBaseTraining2.java (revision 20)
@@ -5,8 +5,18 @@
import de.ugoe.cs.util.console.Console;
+
import weka.core.OptionHandler;
import weka.classifiers.Classifier;
import weka.classifiers.meta.CVParameterSelection;
+/**
+ * WekaBaseTraining2
+ *
+ * Allows specification of the Weka classifier and its params in the XML experiment configuration.
+ *
+ * Important conventions of the XML format:
+ * Cross Validation params come always last and are prepended with -CVPARAM
+ * Example:
+ */
public abstract class WekaBaseTraining2 implements WekaCompatibleTrainer {
@@ -20,13 +30,13 @@
String[] params = parameters.split(" ");
- // first is classifierName
+ // first part of the params is the classifierName (e.g. SMORBF)
classifierName = params[0];
- // all following parameters can be copied from weka!
+ // the following parameters can be copied from weka!
- // second param is classifierClassName
+ // second param is classifierClassName (e.g. weka.classifiers.functions.SMO)
classifierClassName = params[1];
- // rest are params to the specified classifier
+ // rest are params to the specified classifier (e.g. -K weka.classifiers.functions.supportVector.RBFKernel)
classifierParams = Arrays.copyOfRange(params, 2, params.length);
@@ -46,5 +56,5 @@
Classifier obj = (Classifier) c.newInstance();
- // Filter -CVPARAM
+ // Filter out -CVPARAM, these are special because they do not belong to the Weka classifier class as parameters
String[] param = Arrays.copyOf(classifierParams, classifierParams.length);
String[] cvparam = {};
@@ -68,5 +78,5 @@
// we have cross val params
- // cant check on cvparam.length may not be initialized
+ // cant check on cvparam.length here, it may not be initialized
if(cv) {
final CVParameterSelection ps = new CVParameterSelection();
Index: trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaClusterTraining2.java
===================================================================
--- trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaClusterTraining2.java (revision 19)
+++ trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaClusterTraining2.java (revision 20)
@@ -23,9 +23,11 @@
* WekaClusterTraining2
*
- * 1. Cluster traindata
- * 2. for each cluster train a classifier with traindata from cluster
- * 3. match testdata instance to a cluster, then classify with classifier from the cluster
+ * Currently supports only EM Clustering.
*
- * XML config:
+ * 1. Cluster training data
+ * 2. for each cluster train a classifier with training data from cluster
+ * 3. match test data instance to a cluster, then classify with classifier from the cluster
+ *
+ * XML configuration:
*
*
@@ -33,7 +35,4 @@
*
*
- *
- * Questions:
- * - how do we configure the clustering params?
*/
public class WekaClusterTraining2 extends WekaBaseTraining2 implements ITrainingStrategy {
@@ -45,5 +44,4 @@
return classifier;
}
-
@Override
@@ -71,5 +69,12 @@
-
+ /**
+ * Helper method that gives us a clean instance copy with
+ * the values of the instancelist of the first parameter.
+ *
+ * @param instancelist with attributes
+ * @param instance with only values
+ * @return copy of the instance
+ */
private Instance createInstance(Instances instances, Instance instance) {
// attributes for feeding instance to classifier
@@ -96,13 +101,13 @@
}
-
@Override
public double classifyInstance(Instance instance) {
double ret = 0;
try {
+ // 1. copy the instance (keep the class attribute)
Instances traindata = ctraindata.get(0);
Instance classInstance = createInstance(traindata, instance);
- // remove class attribute before clustering
+ // 2. remove class attribute before clustering
Remove filter = new Remove();
filter.setAttributeIndices("" + (traindata.classIndex() + 1));
@@ -110,12 +115,11 @@
traindata = Filter.useFilter(traindata, filter);
+ // 3. copy the instance (without the class attribute) for clustering
Instance clusterInstance = createInstance(traindata, instance);
- // 1. classify testdata instance to a cluster number
+ // 4. match instance without class attribute to a cluster number
int cnum = clusterer.clusterInstance(clusterInstance);
- //Console.traceln(Level.INFO, String.format("instance is in cluster: " + cnum));
-
- // 2. classify testata instance to the classifier
+ // 5. classify instance with class attribute to the classifier of that cluster number
ret = cclassifier.get(cnum).classifyInstance(classInstance);
@@ -127,10 +131,8 @@
}
-
-
@Override
public void buildClassifier(Instances traindata) throws Exception {
- // 1. copy traindata
+ // 1. copy training data
Instances train = new Instances(traindata);
@@ -141,33 +143,28 @@
train = Filter.useFilter(train, filter);
- // 3. cluster data
- //Console.traceln(Level.INFO, String.format("starting clustering"));
-
+ // new objects
cclassifier = new HashMap();
ctraindata = new HashMap();
+ // 3. cluster data
// use standard params for now
clusterer = new EM();
+ // we can set options like so:
//String[] params = {"-N", "100"};
//clusterer.setOptions(params);
+
+ // set max num of clusters to train data size (although we do not want that)
+ clusterer.setMaximumNumberOfClusters(train.size());
+
+ // build clusterer
clusterer.buildClusterer(train);
- // set max num to traindata size
- clusterer.setMaximumNumberOfClusters(train.size());
-
- // 4. get cluster membership of our traindata
- //AddCluster cfilter = new AddCluster();
- //cfilter.setClusterer(clusterer);
- //cfilter.setInputFormat(train);
- //Instances ctrain = Filter.useFilter(train, cfilter);
Instances ctrain = new Instances(train);
- // get traindata per cluster
+ // get train data per cluster
int cnumber;
for ( int j=0; j < ctrain.numInstances(); j++ ) {
- // get the cluster number from the attributes, subract 1 because if we clusterInstance we get 0-n, and this is 1-n
- //cnumber = Integer.parseInt(ctrain.get(j).stringValue(ctrain.get(j).numAttributes()-1).replace("cluster", "")) - 1;
+ cnumber = clusterer.clusterInstance(ctrain.get(j));
- cnumber = clusterer.clusterInstance(ctrain.get(j));
// add training data to list of instances for this cluster number
if ( !ctraindata.containsKey(cnumber) ) {
@@ -178,8 +175,5 @@
}
- // Debug output
- //Console.traceln(Level.INFO, String.format("number of clusters: " + clusterer.numberOfClusters()));
-
- // train one classifier per cluster, we get the clusternumber from the traindata
+ // train one classifier per cluster, we get the cluster number from the training data
Iterator clusternumber = ctraindata.keySet().iterator();
while ( clusternumber.hasNext() ) {
@@ -188,5 +182,5 @@
cclassifier.get(cnumber).buildClassifier(ctraindata.get(cnumber));
- //Console.traceln(Level.INFO, String.format("building classifier in cluster "+cnumber + " with " + ctraindata.get(cnumber).size() + " traindata instances"));
+ //Console.traceln(Level.INFO, String.format("classifier in cluster "+cnumber));
}
}
Index: trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaLocalTraining2.java
===================================================================
--- trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaLocalTraining2.java (revision 19)
+++ trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaLocalTraining2.java (revision 20)
@@ -84,7 +84,11 @@
private HashMap> csize;
+ /* debug vars */
+ @SuppressWarnings("unused")
private boolean show_biggest = true;
+ @SuppressWarnings("unused")
private int CFOUND = 0;
+ @SuppressWarnings("unused")
private int CNOTFOUND = 0;
@@ -260,5 +264,5 @@
//}
- // now it can happen that we dont find a cluster because we deleted it previously (too few instances)
+ // now it can happen that we do not find a cluster because we deleted it previously (too few instances)
// or we get bigger distance measures from weka so that we are completely outside of our clusters.
// in these cases we just find the nearest cluster to our instance and use it for classification.
@@ -280,5 +284,5 @@
}
- // here we have the cluster where an instance has the minimum distance between itself the
+ // here we have the cluster where an instance has the minimum distance between itself and the
// instance we want to classify
// if we still have not found a cluster we exit because something is really wrong
@@ -436,5 +440,5 @@
*/
- // train one classifier per cluster, we get the clusternumber from the traindata
+ // train one classifier per cluster, we get the cluster number from the traindata
int cnumber;
Iterator clusternumber = ctraindata.keySet().iterator();
@@ -444,5 +448,5 @@
while ( clusternumber.hasNext() ) {
cnumber = clusternumber.next();
- cclassifier.put(cnumber,setupClassifier()); // das hier ist der eigentliche trainer
+ cclassifier.put(cnumber,setupClassifier()); // this is the classifier used for the cluster
cclassifier.get(cnumber).buildClassifier(ctraindata.get(cnumber));
//Console.traceln(Level.INFO, String.format("classifier in cluster "+cnumber));
Index: trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaTraining2.java
===================================================================
--- trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaTraining2.java (revision 19)
+++ trunk/CrossPare/src/de/ugoe/cs/cpdp/training/WekaTraining2.java (revision 20)
@@ -10,5 +10,5 @@
/**
- * Programmatic WekaBaggingTraining
+ * Programmatic WekaTraining
*
* first parameter is Trainer Name.