Preliminar changes towards refering to class labels etc. more specifically

Change default label name from Subjet to Label (since label is a more general term). Use different default variable names for classification (label), regression (regressor/regressand), and clustering (ClusterID) Update some (far from all) transforms to accept arguments specifying their input/output variables. Update eval classification to optionally take target variable names as arguments

Preliminar changes towards refering to class labels etc. more specifically
Change default label name from Subjet to Label (since label is a more general term). Use different default variable names for classification (label), regression (regressor/regressand), and clustering (ClusterID) Update some (far from all) transforms to accept arguments specifying their input/output variables. Update eval classification to optionally take target variable names as arguments
Charles Otto
1 parent aebd450c
Showing 22 changed files with 104 additions and 55 deletions
app/br/br.cpp
app/examples/age_estimation.cpp
app/examples/gender_estimation.cpp
openbr/core/bee.cpp
openbr/core/classify.cpp
openbr/core/classify.h
openbr/core/cluster.cpp
openbr/frvt2012.cpp
openbr/gui/classifier.cpp
openbr/openbr.cpp
openbr/openbr.h
openbr/openbr_plugin.cpp
openbr/plugins/cluster.cpp
openbr/plugins/eigen3.cpp
openbr/plugins/gallery.cpp
openbr/plugins/independent.cpp
openbr/plugins/normalize.cpp
openbr/plugins/output.cpp
openbr/plugins/quality.cpp
openbr/plugins/quantize.cpp
@@ -130,7 +130,7 @@ public:
                 br_convert(parv[0], parv[1], parv[2]);
             } else if (!strcmp(fun, "evalClassification")) {
                 check(parc == 2, "Incorrect parameter count for 'evalClassification'.");
-                br_eval_classification(parv[0], parv[1]);
+                br_eval_classification(parv[0], parv[1], parc >= 3 ? parv[2] : NULL, parc >= 4 ? parv[3] : NULL);
             } else if (!strcmp(fun, "evalRegression")) {
                 check(parc == 2, "Incorrect parameter count for 'evalRegression'.");
                 br_eval_regression(parv[0], parv[1]);
@@ -29,7 +29,8 @@
 static void printTemplate(const br::Template &t)
 {
-    printf("%s age: %d\n", qPrintable(t.file.fileName()), int(t.file.get<float>("Subject")));
+    // may use age directly -cao
+    printf("%s age: %d\n", qPrintable(t.file.fileName()), int(t.file.get<float>("Regressand")));
 }
 int main(int argc, char *argv[])
@@ -29,7 +29,8 @@
 static void printTemplate(const br::Template &t)
 {
-    printf("%s gender: %s\n", qPrintable(t.file.fileName()), qPrintable(t.file.get<QString>("Subject")));
+    // may use gender directly -cao
+    printf("%s gender: %s\n", qPrintable(t.file.fileName()), qPrintable(t.file.get<QString>("Label")));
 }
 int main(int argc, char *argv[])
@@ -93,10 +93,10 @@ void BEE::writeSigset(const QString &amp;sigset, const br::FileList &amp;files, bool ign
         QStringList metadata;
         if (!ignoreMetadata)
             foreach (const QString &key, file.localKeys()) {
-                if ((key == "Index") || (key == "Subject")) continue;
+                if ((key == "Index") || (key == "Label")) continue;
                 metadata.append(key+"=\""+QtUtils::toString(file.value(key))+"\"");
             }
-        lines.append("\t<biometric-signature name=\"" + file.get<QString>("Subject") +"\">");
+        lines.append("\t<biometric-signature name=\"" + file.get<QString>("Label") +"\">");
         lines.append("\t\t<presentation file-name=\"" + file.name + "\" " + metadata.join(" ") + "/>");
         lines.append("\t</biometric-signature>");
     }
@@ -260,10 +260,10 @@ void BEE::makeMask(const QString &amp;targetInput, const QString &amp;queryInput, const 
 cv::Mat BEE::makeMask(const br::FileList &targets, const br::FileList &queries, int partition)
 {
-    // Would like to use indexProperty for this, but didn't make a version of that for Filelist yet
-    // -cao
-    QList<QString> targetLabels = targets.get<QString>("Subject", "-1");
-    QList<QString> queryLabels = queries.get<QString>("Subject", "-1");
+    // Direct use of "Label" isn't general, also would prefer to use indexProperty, rather than
+    // doing string comparisons (but that isn't implemented yet for FileList) -cao
+    QList<QString> targetLabels = targets.get<QString>("Label", "-1");
+    QList<QString> queryLabels = queries.get<QString>("Label", "-1");
     QList<int> targetPartitions = targets.crossValidationPartitions();
     QList<int> queryPartitions = queries.crossValidationPartitions();
@@ -31,7 +31,7 @@ struct Counter
     }
 };
-void br::EvalClassification(const QString &predictedInput, const QString &truthInput)
+void br::EvalClassification(const QString &predictedInput, const QString &truthInput, const QString & predictedProperty, const QString & truthProperty)
 {
     qDebug("Evaluating classification of %s against %s", qPrintable(predictedInput), qPrintable(truthInput));
@@ -44,9 +44,8 @@ void br::EvalClassification(const QString &amp;predictedInput, const QString &amp;truthI
         if (predicted[i].file.name != truth[i].file.name)
             qFatal("Input order mismatch.");
-        // Typically these lists will be of length one, but this generalization allows measuring multi-class labeling accuracy.
-        QString predictedSubject = predicted[i].file.get<QString>("Subject");
-        QString trueSubject = truth[i].file.get<QString>("Subject");
+        QString predictedSubject = predicted[i].file.get<QString>(predictedProperty);
+        QString trueSubject = truth[i].file.get<QString>(truthProperty);
         QStringList predictedSubjects(predictedSubject);
         QStringList trueSubjects(trueSubject);
@@ -99,13 +98,19 @@ void br::EvalRegression(const QString &amp;predictedInput, const QString &amp;truthInput
     if (predicted.size() != truth.size()) qFatal("Input size mismatch.");
     float rmsError = 0;
+    float maeError = 0;
+    // Direct use of Regressor/Regressand is not general -cao
     QStringList truthValues, predictedValues;
     for (int i=0; i<predicted.size(); i++) {
         if (predicted[i].file.name != truth[i].file.name)
             qFatal("Input order mismatch.");
-        rmsError += pow(predicted[i].file.get<float>("Subject")-truth[i].file.get<float>("Subject"), 2.f);
-        truthValues.append(QString::number(truth[i].file.get<float>("Subject")));
-        predictedValues.append(QString::number(predicted[i].file.get<float>("Subject")));
+
+        float difference = predicted[i].file.get<float>("Regressand") - truth[i].file.get<float>("Regressor");
+
+        rmsError += pow(difference, 2.f);
+        maeError += fabsf(difference);
+        truthValues.append(QString::number(truth[i].file.get<float>("Regressor")));
+        predictedValues.append(QString::number(predicted[i].file.get<float>("Regressand")));
     }
     QStringList rSource;
@@ -125,4 +130,6 @@ void br::EvalRegression(const QString &amp;predictedInput, const QString &amp;truthInput
     if (success) QtUtils::showFile("EvalRegression.pdf");
     qDebug("RMS Error = %f", sqrt(rmsError/predicted.size()));
+    qDebug("MAE = %f", maeError/predicted.size());
+
 }
@@ -22,7 +22,7 @@
 namespace br
 {
-    void EvalClassification(const QString &predictedInput, const QString &truthInput);
+    void EvalClassification(const QString &predictedInput, const QString &truthInput, const QString & predictedProperty="Label", const QString & truthProperty="Label");
     void EvalRegression(const QString &predictedInput, const QString &truthInput);
 }
@@ -279,8 +279,8 @@ void br::EvalClustering(const QString &amp;csv, const QString &amp;input)
     qDebug("Evaluating %s against %s", qPrintable(csv), qPrintable(input));
     // We assume clustering algorithms store assigned cluster labels as integers (since the clusters are
-    // not named).
-    QList<int> labels = TemplateList::fromGallery(input).files().get<int>("Subject");
+    // not named). Direct use of ClusterID is not general -cao
+    QList<int> labels = TemplateList::fromGallery(input).files().get<int>("ClusterID");
     QHash<int, int> labelToIndex;
     int nClusters = 0;
@@ -132,7 +132,8 @@ int32_t SdkEstimator::estimate_age(const ONEFACE &amp;input_face, int32_t &amp;age)
     TemplateList templates;
     templates.append(templateFromONEFACE(input_face));
     templates >> *frvt2012_age_transform.data();
-    age = templates.first().file.get<float>("Subject");
+    // should maybe use "Age" directly -cao
+    age = templates.first().file.get<float>("Regressand");
     return templates.first().file.failed() ? 4 : 0;
 }
@@ -141,6 +142,7 @@ int32_t SdkEstimator::estimate_gender(const ONEFACE &amp;input_face, int8_t &amp;gender,
     TemplateList templates;
     templates.append(templateFromONEFACE(input_face));
     templates >> *frvt2012_gender_transform.data();
-    mf = gender = templates.first().file.get<QString>("Subject") == "Male" ? 0 : 1;
+    // Should maybe use "Gender" directly -cao
+    mf = gender = templates.first().file.get<QString>("Label") == "Male" ? 0 : 1;
     return templates.first().file.failed() ? 4 : 0;
 }
@@ -43,14 +43,16 @@ void Classifier::_classify(File file)
             continue;
         if (algorithm == "GenderClassification") {
+            // Should maybe use gender directly -cao
             key = "Gender";
-            value = (f.get<QString>("Subject"));
+            value = f.get<QString>("Label");
         } else if (algorithm == "AgeRegression") {
             key = "Age";
-            value = QString::number(int(f.get<float>("Subject")+0.5)) + " Years";
+            // similarly, age -cao
+            value = QString::number(int(f.get<float>("Regressand")+0.5)) + " Years";
         } else {
             key = algorithm;
-            value = f.get<QString>("Subject");
+            value = f.get<QString>("Label");
         }
         break;
     }
@@ -77,9 +77,14 @@ float br_eval(const char *simmat, const char *mask, const char *csv)
     return Evaluate(simmat, mask, csv);
 }
-void br_eval_classification(const char *predicted_input, const char *truth_input)
-{
-    EvalClassification(predicted_input, truth_input);
+void br_eval_classification(const char *predicted_input, const char *truth_input, const char *predicted_property, const char * truth_property)
+{
+    if (predicted_property && truth_property)
+        EvalClassification(predicted_input, truth_input, predicted_property, truth_property);
+    else if (predicted_property)
+        EvalClassification(predicted_input, truth_input, predicted_property);
+    else
+        EvalClassification(predicted_input, truth_input);
 }
 void br_eval_clustering(const char *csv, const char *input)
@@ -164,7 +164,7 @@ BR_EXPORT float br_eval(const char *simmat, const char *mask, const char *csv = 
  * \param truth_input The ground truth br::Input.
  * \see br_enroll
  */
-BR_EXPORT void br_eval_classification(const char *predicted_input, const char *truth_input);
+BR_EXPORT void br_eval_classification(const char *predicted_input, const char *truth_input, const char * predicted_property, const char * truth_property);
 /*!
  * \brief Evaluates and prints clustering accuracy to the terminal.
@@ -412,6 +412,7 @@ TemplateList TemplateList::fromGallery(const br::File &amp;gallery)
                     // of target images to every partition
                     newTemplates[i].file.set("Partition", -1);
                 } else {
+                    // Direct use of "Subject" is not general -cao
                     const QByteArray md5 = QCryptographicHash::hash(newTemplates[i].file.get<QString>("Subject").toLatin1(), QCryptographicHash::Md5);
                     // Select the right 8 hex characters so that it can be represented as a 64 bit integer without overflow
                     newTemplates[i].file.set("Partition", md5.toHex().right(8).toULongLong(0, 16) % crossValidate);
@@ -89,10 +89,12 @@ class KNNTransform : public Transform
     Q_PROPERTY(br::Distance *distance READ get_distance WRITE set_distance RESET reset_distance STORED false)
     Q_PROPERTY(bool weighted READ get_weighted WRITE set_weighted RESET reset_weighted STORED false)
     Q_PROPERTY(int numSubjects READ get_numSubjects WRITE set_numSubjects RESET reset_numSubjects STORED false)
+    Q_PROPERTY(QString inputVariable READ get_inputVariable WRITE set_inputVariable RESET reset_inputVariable STORED false)
     BR_PROPERTY(int, k, 1)
     BR_PROPERTY(br::Distance*, distance, NULL)
     BR_PROPERTY(bool, weighted, false)
     BR_PROPERTY(int, numSubjects, 1)
+    BR_PROPERTY(QString, inputVariable, "Label")
     TemplateList gallery;
@@ -111,13 +113,13 @@ class KNNTransform : public Transform
             QHash<QString, float> votes;
             const int max = (k < 1) ? sortedScores.size() : std::min(k, sortedScores.size());
             for (int j=0; j<max; j++)
-                votes[gallery[sortedScores[j].second].file.get<QString>("Subject")] += (weighted ? sortedScores[j].first : 1);
+                votes[gallery[sortedScores[j].second].file.get<QString>(inputVariable)] += (weighted ? sortedScores[j].first : 1);
             subjects.append(votes.keys()[votes.values().indexOf(Common::Max(votes.values()))]);
             // Remove subject from consideration
             if (subjects.size() < numSubjects)
                 for (int j=sortedScores.size()-1; j>=0; j--)
-                    if (gallery[sortedScores[j].second].file.get<QString>("Subject") == subjects.last())
+                    if (gallery[sortedScores[j].second].file.get<QString>(inputVariable) == subjects.last())
                         sortedScores.removeAt(j);
         }
@@ -318,10 +318,12 @@ class LDATransform : public Transform
     Q_PROPERTY(bool pcaWhiten READ get_pcaWhiten WRITE set_pcaWhiten RESET reset_pcaWhiten STORED false)
     Q_PROPERTY(int directLDA READ get_directLDA WRITE set_directLDA RESET reset_directLDA STORED false)
     Q_PROPERTY(float directDrop READ get_directDrop WRITE set_directDrop RESET reset_directDrop STORED false)
+    Q_PROPERTY(QString inputVariable READ get_inputVariable WRITE set_inputVariable RESET reset_inputVariable STORED false)
     BR_PROPERTY(float, pcaKeep, 0.98)
     BR_PROPERTY(bool, pcaWhiten, false)
     BR_PROPERTY(int, directLDA, 0)
     BR_PROPERTY(float, directDrop, 0.1)
+    BR_PROPERTY(QString, inputVariable, "Label")
     int dimsOut;
     Eigen::VectorXf mean;
@@ -330,7 +332,7 @@ class LDATransform : public Transform
     void train(const TemplateList &_trainingSet)
     {
         // creates "Label"
-        TemplateList trainingSet = TemplateList::relabel(_trainingSet, "Subject");
+        TemplateList trainingSet = TemplateList::relabel(_trainingSet, inputVariable);
         int instances = trainingSet.size();
@@ -71,7 +71,7 @@ class arffGallery : public Gallery
         }
         arffFile.write(qPrintable(OpenCVUtils::matrixToStringList(t).join(',')));
-        arffFile.write(qPrintable(",'" + t.file.get<QString>("Subject") + "'\n"));
+        arffFile.write(qPrintable(",'" + t.file.get<QString>("Label") + "'\n"));
     }
 };
@@ -874,7 +874,7 @@ class statGallery : public Gallery
     void write(const Template &t)
     {
-        subjects.insert(t.file.get<QString>("Subject"));
+        subjects.insert(t.file.get<QString>("Label"));
         bytes.append(t.bytes());
     }
 };
@@ -20,11 +20,11 @@ static TemplateList Downsample(const TemplateList &amp;templates, const Transform *t
     const bool atLeast = transform->instances < 0;
     const int instances = abs(transform->instances);
-    QList<QString> allLabels = templates.get<QString>("Subject");
+    QList<QString> allLabels = templates.get<QString>("Label");
     QList<QString> uniqueLabels = allLabels.toSet().toList();
     qSort(uniqueLabels);
-    QMap<QString,int> counts = templates.countValues<QString>("Subject", instances != std::numeric_limits<int>::max());
+    QMap<QString,int> counts = templates.countValues<QString>("Label", instances != std::numeric_limits<int>::max());
     if ((instances != std::numeric_limits<int>::max()) && (transform->classes != std::numeric_limits<int>::max()))
         foreach (const QString & label, counts.keys())
@@ -127,7 +127,7 @@ private:
         Mat m;
         OpenCVUtils::toMat(data.data()).convertTo(m, CV_64F);
-        const QList<int> labels = data.indexProperty("Subject");
+        const QList<int> labels = data.indexProperty("Label");
         const int dims = m.cols;
         vector<Mat> mv, av, bv;
@@ -146,8 +146,8 @@ class meltOutput : public MatrixOutput
         QStringList lines;
         if (file.baseName() != "terminal") lines.append(QString("Query,Target,Mask,Similarity%1").arg(keys));
-        QList<QString> queryLabels = queryFiles.get<QString>("Subject");
-        QList<QString> targetLabels = targetFiles.get<QString>("Subject");
+        QList<QString> queryLabels = queryFiles.get<QString>("Label");
+        QList<QString> targetLabels = targetFiles.get<QString>("Label");
         for (int i=0; i<queryFiles.size(); i++) {
             for (int j=(selfSimilar ? i+1 : 0); j<targetFiles.size(); j++) {
@@ -298,7 +298,7 @@ class txtOutput : public MatrixOutput
         if (file.isNull() || targetFiles.isEmpty() || queryFiles.isEmpty()) return;
         QStringList lines;
         foreach (const File &file, queryFiles)
-            lines.append(file.name + " " + file.get<QString>("Subject"));
+            lines.append(file.name + " " + file.get<QString>("Label"));
         QtUtils::writeFile(file, lines);
     }
 };
@@ -426,7 +426,7 @@ class rankOutput : public MatrixOutput
             typedef QPair<float,int> Pair;
             int rank = 1;
             foreach (const Pair &pair, Common::Sort(OpenCVUtils::matrixToVector<float>(data.row(i)), true)) {
-                if (targetFiles[pair.second].get<QString>("Subject") == queryFiles[i].get<QString>("Subject")) {
+                if (targetFiles[pair.second].get<QString>("Label") == queryFiles[i].get<QString>("Label")) {
                     ranks.append(rank);
                     positions.append(pair.second);
                     scores.append(pair.first);
@@ -26,10 +26,10 @@ class ImpostorUniquenessMeasureTransform : public Transform
     float calculateIUM(const Template &probe, const TemplateList &gallery) const
     {
-        const QString probeLabel = probe.file.get<QString>("Subject");
+        const QString probeLabel = probe.file.get<QString>("Label");
         TemplateList subset = gallery;
         for (int j=subset.size()-1; j>=0; j--)
-            if (subset[j].file.get<QString>("Subject") == probeLabel)
+            if (subset[j].file.get<QString>("Label") == probeLabel)
                 subset.removeAt(j);
         QList<float> scores = distance->compare(subset, probe);
@@ -158,7 +158,7 @@ class MatchProbabilityDistance : public Distance
     {
         distance->train(src);
-        const QList<int> labels = src.indexProperty("Subject");
+        const QList<int> labels = src.indexProperty("Label");
         QScopedPointer<MatrixOutput> matrixOutput(MatrixOutput::make(FileList(src.size()), FileList(src.size())));
         distance->compare(src, src, matrixOutput.data());
@@ -228,7 +228,7 @@ class HeatMapDistance : public Distance
     {
         distance->train(src);
-        const QList<int> labels = src.indexProperty("Subject");
+        const QList<int> labels = src.indexProperty("Label");
         QList<TemplateList> patches;
@@ -316,7 +316,7 @@ class UnitDistance : public Distance
     void train(const TemplateList &templates)
     {
         const TemplateList samples = templates.mid(0, 2000);
-        const QList<int> sampleLabels = samples.indexProperty("Subject");
+        const QList<int> sampleLabels = samples.indexProperty("Label");
         QScopedPointer<MatrixOutput> matrixOutput(MatrixOutput::make(FileList(samples.size()), FileList(samples.size())));
         Distance::compare(samples, samples, matrixOutput.data());
@@ -150,7 +150,7 @@ class BayesianQuantizationDistance : public Distance
             qFatal("Expected sigle matrix templates of type CV_8UC1!");
         const Mat data = OpenCVUtils::toMat(src.data());
-        const QList<int> templateLabels = src.indexProperty("Subject");
+        const QList<int> templateLabels = src.indexProperty("Label");
         loglikelihoods = QVector<float>(data.cols*256, 0);
         QFutureSynchronizer<void> futures;
@@ -474,7 +474,7 @@ private:
         Mat data = OpenCVUtils::toMat(src.data());
         const int step = getStep(data.cols);
-        const QList<int> labels = src.indexProperty("Subject");
+        const QList<int> labels = src.indexProperty("Label");
         Mat &lut = ProductQuantizationLUTs[index];
         lut = Mat(getDims(data.cols), 256*(256+1)/2, CV_32FC1);
@@ -77,7 +77,7 @@ class BayesianQuantizationTransform : public Transform
     void train(const TemplateList &src)
     {
         const Mat data = OpenCVUtils::toMat(src.data());
-        const QList<int> labels = src.indexProperty("Subject");
+        const QList<int> labels = src.indexProperty("Label");
         thresholds = QVector<float>(256*data.cols);
@@ -101,6 +101,8 @@ class SVMTransform : public Transform
     Q_PROPERTY(Type type READ get_type WRITE set_type RESET reset_type STORED false)
     Q_PROPERTY(float C READ get_C WRITE set_C RESET reset_C STORED false)
     Q_PROPERTY(float gamma READ get_gamma WRITE set_gamma RESET reset_gamma STORED false)
+    Q_PROPERTY(QString inputVariable READ get_inputVariable WRITE set_inputVariable RESET reset_inputVariable STORED false)
+    Q_PROPERTY(QString outputVariable READ get_outputVariable WRITE set_outputVariable RESET reset_outputVariable STORED false)
 public:
     enum Kernel { Linear = CvSVM::LINEAR,
@@ -119,6 +121,9 @@ private:
     BR_PROPERTY(Type, type, C_SVC)
     BR_PROPERTY(float, C, -1)
     BR_PROPERTY(float, gamma, -1)
+    BR_PROPERTY(QString, inputVariable, "")
+    BR_PROPERTY(QString, outputVariable, "")
+
     SVM svm;
     QHash<QString, int> labelMap;
@@ -128,14 +133,14 @@ private:
     {
         Mat data = OpenCVUtils::toMat(_data.data());
         Mat lab;
-        // If we are doing regression, assume subject has float values
+        // If we are doing regression, the input variable should have float values
         if (type == EPS_SVR || type == NU_SVR) {
-            lab = OpenCVUtils::toMat(_data.get<float>("Subject"));
+            lab = OpenCVUtils::toMat(_data.get<float>(inputVariable));
         }
-        // If we are doing classification, assume subject has discrete values, map them
+        // If we are doing classification, we should be dealing with discrete values. Map them
         // and store the mapping data
         else {
-            QList<int> dataLabels = _data.indexProperty("Subject", labelMap, reverseLookup);
+            QList<int> dataLabels = _data.indexProperty(inputVariable, labelMap, reverseLookup);
             lab = OpenCVUtils::toMat(dataLabels);
         }
         trainSVM(svm, data, lab, kernel, type, C, gamma);
@@ -146,9 +151,9 @@ private:
         dst = src;
         float prediction = svm.predict(src.m().reshape(1, 1));
         if (type == EPS_SVR || type == NU_SVR)
-            dst.file.set("Subject", prediction);
+            dst.file.set(outputVariable, prediction);
         else
-            dst.file.set("Subject", reverseLookup[prediction]);
+            dst.file.set(outputVariable, reverseLookup[prediction]);
     }
     void store(QDataStream &stream) const
@@ -162,6 +167,24 @@ private:
         loadSVM(svm, stream);
         stream >> labelMap >> reverseLookup;
     }
+
+    void init()
+    {
+        // Since SVM can do regression or classification, we have to check the problem type before
+        // specifying target variable names
+        if (inputVariable.isEmpty())
+        {
+            if (type == EPS_SVR || type == NU_SVR) {
+                inputVariable = "Regressor";
+                if (outputVariable.isEmpty())
+                    outputVariable = "Regressand";
+            }
+            else
+                inputVariable = "Label";
+        }
+        if (outputVariable.isEmpty())
+            outputVariable = inputVariable;
+    }
 };
 BR_REGISTER(Transform, SVMTransform)
@@ -178,6 +201,8 @@ class SVMDistance : public Distance
     Q_ENUMS(Type)
     Q_PROPERTY(Kernel kernel READ get_kernel WRITE set_kernel RESET reset_kernel STORED false)
     Q_PROPERTY(Type type READ get_type WRITE set_type RESET reset_type STORED false)
+    Q_PROPERTY(QString inputVariable READ get_inputVariable WRITE set_inputVariable RESET reset_inputVariable STORED false)
+
 public:
     enum Kernel { Linear = CvSVM::LINEAR,
@@ -194,13 +219,14 @@ public:
 private:
     BR_PROPERTY(Kernel, kernel, Linear)
     BR_PROPERTY(Type, type, EPS_SVR)
+    BR_PROPERTY(QString, inputVariable, "Label")
     SVM svm;
     void train(const TemplateList &src)
     {
         const Mat data = OpenCVUtils::toMat(src.data());
-        const QList<int> lab = src.indexProperty("Subject");
+        const QList<int> lab = src.indexProperty(inputVariable);
         const int instances = data.rows * (data.rows+1) / 2;
         Mat deltaData(instances, data.cols, data.type());