CrossValidation with same training across all splits

Scott Klum
1 parent d29e2df1
Showing 8 changed files with 33 additions and 11 deletions
openbr/core/core.cpp
openbr/core/plot.cpp
openbr/openbr_plugin.cpp
openbr/openbr_plugin.h
openbr/plugins/eigen3.cpp
openbr/plugins/format.cpp
openbr/plugins/stasm.cpp
openbr/plugins/validate.cpp
@@ -248,9 +248,8 @@ struct AlgorithmCore
                     if (!partitionSizes.empty()) targetPartitions = targets.partition(partitionSizes);
                     else targetPartitions.append(targets);
  
-                    if (queryPartitions[i].first().size() != targetPartitions[i].first().size()) qFatal("Query and target templates have different number of matrices.");
-
                     outputs[i]->setBlock(queryBlock, targetBlock);
+
                     distance->compare(targetPartitions[i], queryPartitions[i], outputs[i]);
  
                     Globals->currentStep += double(targets.size()) * double(queries.size());
@@ -268,7 +268,7 @@ float Evaluate(const Mat &amp;simmat, const Mat &amp;mask, const QString &amp;csv)
     }
  
     // Write Cumulative Match Characteristic (CMC) curve
-    const int Max_Retrieval = 100;
+    const int Max_Retrieval = 200;
     const int Report_Retrieval = 5;
  
     float reportRetrievalRate = -1;
@@ -472,6 +472,8 @@ struct RPlot
     }
 };
  
+// Does not work if dataset folder starts with a number
+
 bool Plot(const QStringList &files, const br::File &destination, bool show)
 {
     qDebug("Plotting %d file(s) to %s", files.size(), qPrintable(destination));
@@ -137,7 +137,7 @@ QVariant File::value(const QString &amp;key) const
  
 QVariant File::parse(const QString &value)
 {
-    bool ok;
+    bool ok = false;
     const QPointF point = QtUtils::toPoint(value, &ok);
     if (ok) return point;
     const QRectF rect = QtUtils::toRect(value, &ok);
@@ -1003,7 +1003,7 @@ void Output::reformat(const FileList &amp;targetFiles, const FileList &amp;queryFiles, c
     const int columns = targetFiles.size();
     for (int i=0; i<rows; i++)
         for (int j=0; j<columns; j++)
-            o->setRelative(m.at<float>(i,i), i, j);
+            o->setRelative(m.at<float>(i,j), i, j);
 }
  
 /* Output - protected methods */
@@ -410,7 +410,7 @@ struct TemplateList : public QList&lt;Template&gt;
             sum+=partitionSizes[i];
         }
  
-        if (sum != first().size()) qFatal("Partition sizes do not span template matrices properly");
+        if (sum != first().size()) qFatal("Partition sizes %i do not span template matrices %i properly", sum, first().size());
  
         foreach (const Template &t, *this) {
             int index = 0;
@@ -330,6 +330,7 @@ class LDATransform : public Transform
     void train(const TemplateList &_trainingSet)
     {
         TemplateList trainingSet = TemplateList::relabel(_trainingSet);
+
         int instances = trainingSet.size();
  
         // Perform PCA dimensionality reduction
@@ -317,6 +317,7 @@ BR_REGISTER(Format, maskFormat)
  * \brief MATLAB <tt>.mat</tt> format.
  * \author Josh Klontz \cite jklontz
  * http://www.mathworks.com/help/pdf_doc/matlab/matfile_format.pdf
+ * \note matFormat is known not to work with compressed matrices
  */
 class matFormat : public Format
 {
@@ -324,12 +325,18 @@ class matFormat : public Format
  
     struct Element
     {
+        // It is always best to cast integers to a Qt integer type, such as qint16 or quint32, when reading and writing.
+        // This ensures that you always know exactly what size integers you are reading and writing, no matter what the
+        // underlying platform and architecture the application happens to be running on.
+        // http://qt-project.org/doc/qt-4.8/datastreamformat.html
         quint32 type, bytes;
         QByteArray data;
         Element() : type(0), bytes(0) {}
         Element(QDataStream &stream)
             : type(0), bytes(0)
         {
+            // Read first 4 bytes into type (32 bit integer),
+            // specifying the type of data used
             if (stream.readRawData((char*)&type, 4) != 4)
                 qFatal("Unexpected end of file.");
  
@@ -340,11 +347,16 @@ class matFormat : public Format
                 bytes = bytes >> 16;
             } else {
                 // Regular format
+                // Read 4 bytes into bytes (32 bit integer),
+                // specifying the size of the element
                 if (stream.readRawData((char*)&bytes, 4) != 4)
                     qFatal("Unexpected end of file.");
             }
  
+            // Set the size of data to bytes
             data.resize(bytes);
+
+            // Read bytes amount of data from the file into data
             if (int(bytes) != stream.readRawData(data.data(), bytes))
                 qFatal("Unexpected end of file.");
  
@@ -372,8 +384,9 @@ class matFormat : public Format
         while (!f.atEnd()) {
             Element element(f);
  
-            // miCOMPRESS
+            // miCOMPRESSED
             if (element.type == 15) {
+                // Prepend the number of bytes to element.data
                 element.data.prepend((char*)&element.bytes, 4); // Qt zlib wrapper requires this to preallocate the buffer
                 QDataStream uncompressed(qUncompress(element.data));
                 element = Element(uncompressed);
@@ -34,7 +34,7 @@ BR_REGISTER(Initializer, StasmInitializer)
  * \author Scott Klum \cite sklum
  */
 // TODO: Use a global mutex to prevent concurrent calls to AsmSearchDll
-#if 0
+
 class StasmTransform : public UntrainableTransform
 {
     Q_OBJECT
@@ -46,6 +46,9 @@ class StasmTransform : public UntrainableTransform
  
     void project(const Template &src, Template &dst) const
     {
+        static QMutex mutex;
+        QMutexLocker locker(&mutex);
+
         int nlandmarks;
         int landmarks[500];
  
@@ -68,7 +71,6 @@ class StasmTransform : public UntrainableTransform
 };
  
 BR_REGISTER(Transform, StasmTransform)
-#endif
  
 } // namespace br
  
@@ -52,8 +52,13 @@ class CrossValidateTransform : public MetaTransform
     }
  
     void project(const Template &src, Template &dst) const
-    {
-        transforms[src.file.get<int>("Partition", 0)]->project(src, dst);
+    {        
+        // If the src partition is greater than the number of training partitions,
+        // assume that projection should be done using the same training data for all partitions.
+        int partition = src.file.get<int>("Partition", 0);
+        if (partition >= transforms.size()-1) partition = 0;
+
+        transforms[partition]->project(src, dst);
     }
  
     void store(QDataStream &stream) const