changeset 25:2c913b88b808

Fix folding

--- a/src/CRP.cpp	Tue Sep 29 16:36:23 2015 +0100
+++ b/src/CRP.cpp	Tue Sep 29 16:51:49 2015 +0100
@@ -24,9 +24,14 @@
     if (in.empty()) {
 	return in;
     }
+
+    // The chroma processing chain requires a 120-bin pitch filterbank
+    // output, even though ours only actually contains 88 bins.
+
+    int size = 120;
+    
     if (!m_dctReduce) {
-	m_size = in[0].size();
-	m_dctReduce = new DCTReduce(m_size, m_params.coefficientsToDrop);
+	m_dctReduce = new DCTReduce(size, m_params.coefficientsToDrop);
     }
 
     RealBlock out;
@@ -37,10 +42,14 @@
             col = LogCompress::process(col, m_params.logFactor, m_params.logAddTerm);
 	}
 
+        RealColumn resized(20, 0.0);
+        resized.insert(resized.end(), col.begin(), col.end());
+        resized.resize(size);
+        
         out.push_back(Normalise::normalise
                       (OctaveFold::process
-                       (m_dctReduce->process(col)),
-                       m_params.normP));
+                       (m_dctReduce->process(resized)),
+                       m_params.normP, m_params.normThresh));
     }        
 
     return out;

--- a/src/CRP.h	Tue Sep 29 16:36:23 2015 +0100
+++ b/src/CRP.h	Tue Sep 29 16:51:49 2015 +0100
@@ -18,23 +18,24 @@
         double logFactor;
         double logAddTerm;
         int normP;
+        double normThresh;
 	Parameters() :
-            coefficientsToDrop(54),
+            coefficientsToDrop(55),
             applyLogCompression(true),
             logFactor(1000.0),
             logAddTerm(1.0),
-            normP(2)
+            normP(2),
+            normThresh(1e-6)
         { }
     };
 	
-    CRP(Parameters params) : m_params(params), m_size(0), m_dctReduce(0) { }
+    CRP(Parameters params) : m_params(params), m_dctReduce(0) { }
     ~CRP();
 
     RealBlock process(const RealBlock &in);
     
 private:
     Parameters m_params;
-    int m_size;
     DCTReduce *m_dctReduce;
 };
 

--- a/src/Normalise.h	Tue Sep 29 16:36:23 2015 +0100
+++ b/src/Normalise.h	Tue Sep 29 16:51:49 2015 +0100
@@ -9,10 +9,10 @@
 {
 public:
     static double norm(std::vector<double> v,
-		       int p = 2); // L^p norm
+		       int p); // L^p norm
 
     static std::vector<double> normalise(std::vector<double> v,
-					 int p = 2,
+					 int p,
 					 double threshold = 1e-6);
 };
 

--- a/src/OctaveFold.h	Tue Sep 29 16:36:23 2015 +0100
+++ b/src/OctaveFold.h	Tue Sep 29 16:51:49 2015 +0100
@@ -8,17 +8,10 @@
 class OctaveFold
 {
 public:
-    /**
-     * Take an 88-bin pitch feature and sum it into a 12-bin
-     * octave. Each bin gets summed into chroma bin corresponding to
-     * its MIDI pitch modulo 12. The first 20 MIDI pitches are
-     * missing, so bin number n (for n from 0 to 87) is for MIDI pitch
-     * 21+n.
-     */
     static std::vector<double> process(std::vector<double> in) {
 	std::vector<double> out(12, 0.0);
 	for (int i = 0; i < int(in.size()); ++i) {
-	    out[(21+i) % 12] += in[i];
+	    out[(i+1) % 12] += in[i];
 	}
 	return std::move(out);
     }

--- a/test/examples/README.txt	Tue Sep 29 16:36:23 2015 +0100
+++ b/test/examples/README.txt	Tue Sep 29 16:51:49 2015 +0100
@@ -39,5 +39,7 @@
 The proposition that these are intended to help test is: If Tipic
 produces the pitch features found in ylsf-30sec-tipic-pitch.csv, and
 if we assume the MATLAB implementations are correct, then Tipic should
-also produce chroma features matching those in these files.
+also produce chroma features matching those in these files. (Up to
+rounding error based on the limited precision with which the original
+pitch values were written out to CSV.)