c@41: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
c@41: 
c@41: /*
c@41:  * SegmenterPlugin.cpp
c@41:  *
c@41:  * Copyright 2008 Centre for Digital Music, Queen Mary, University of London.
c@41:  * All rights reserved.
c@41:  */
c@41: 
c@41: #include <iostream>
c@44: #include <cstdio>
c@41: 
c@41: #include "SimilarityPlugin.h"
c@42: #include "base/Pitch.h"
c@41: #include "dsp/mfcc/MFCC.h"
c@42: #include "dsp/chromagram/Chromagram.h"
c@41: #include "dsp/rateconversion/Decimator.h"
c@41: 
c@41: using std::string;
c@41: using std::vector;
c@41: using std::cerr;
c@41: using std::endl;
c@41: using std::ostringstream;
c@41: 
c@41: SimilarityPlugin::SimilarityPlugin(float inputSampleRate) :
c@41:     Plugin(inputSampleRate),
c@42:     m_type(TypeMFCC),
c@41:     m_mfcc(0),
c@42:     m_chromagram(0),
c@41:     m_decimator(0),
c@42:     m_featureColumnSize(20),
c@41:     m_blockSize(0),
c@41:     m_channels(0)
c@41: {
c@41: 	
c@41: }
c@41: 
c@41: SimilarityPlugin::~SimilarityPlugin()
c@41: {
c@41:     delete m_mfcc;
c@42:     delete m_chromagram;
c@41:     delete m_decimator;
c@41: }
c@41: 
c@41: string
c@41: SimilarityPlugin::getIdentifier() const
c@41: {
c@41:     return "qm-similarity";
c@41: }
c@41: 
c@41: string
c@41: SimilarityPlugin::getName() const
c@41: {
c@41:     return "Similarity";
c@41: }
c@41: 
c@41: string
c@41: SimilarityPlugin::getDescription() const
c@41: {
c@42:     return "Return a distance matrix for similarity between the input audio channels";
c@41: }
c@41: 
c@41: string
c@41: SimilarityPlugin::getMaker() const
c@41: {
c@41:     return "Chris Cannam, Queen Mary, University of London";
c@41: }
c@41: 
c@41: int
c@41: SimilarityPlugin::getPluginVersion() const
c@41: {
c@41:     return 1;
c@41: }
c@41: 
c@41: string
c@41: SimilarityPlugin::getCopyright() const
c@41: {
c@41:     return "Copyright (c) 2008 - All Rights Reserved";
c@41: }
c@41: 
c@41: size_t
c@41: SimilarityPlugin::getMinChannelCount() const
c@41: {
c@43:     return 1;
c@41: }
c@41: 
c@41: size_t
c@41: SimilarityPlugin::getMaxChannelCount() const
c@41: {
c@41:     return 1024;
c@44: //    return 1;
c@41: }
c@41: 
c@41: bool
c@41: SimilarityPlugin::initialise(size_t channels, size_t stepSize, size_t blockSize)
c@41: {
c@41:     if (channels < getMinChannelCount() ||
c@41: 	channels > getMaxChannelCount()) return false;
c@41: 
c@41:     if (stepSize != getPreferredStepSize()) {
c@43:         //!!! actually this perhaps shouldn't be an error... similarly
c@43:         //using more than getMaxChannelCount channels
c@41:         std::cerr << "SimilarityPlugin::initialise: supplied step size "
c@41:                   << stepSize << " differs from required step size "
c@41:                   << getPreferredStepSize() << std::endl;
c@41:         return false;
c@41:     }
c@41: 
c@41:     if (blockSize != getPreferredBlockSize()) {
c@41:         std::cerr << "SimilarityPlugin::initialise: supplied block size "
c@41:                   << blockSize << " differs from required block size "
c@41:                   << getPreferredBlockSize() << std::endl;
c@41:         return false;
c@41:     }        
c@41:     
c@41:     m_blockSize = blockSize;
c@41:     m_channels = channels;
c@41: 
c@44:     m_lastNonEmptyFrame = std::vector<int>(m_channels);
c@44:     for (int i = 0; i < m_channels; ++i) m_lastNonEmptyFrame[i] = -1;
c@44:     m_frameNo = 0;
c@44: 
c@41:     int decimationFactor = getDecimationFactor();
c@41:     if (decimationFactor > 1) {
c@42:         m_decimator = new Decimator(m_blockSize, decimationFactor);
c@41:     }
c@41: 
c@42:     if (m_type == TypeMFCC) {
c@42: 
c@42:         m_featureColumnSize = 20;
c@42: 
c@42:         MFCCConfig config;
c@42:         config.FS = lrintf(m_inputSampleRate) / decimationFactor;
c@42:         config.fftsize = 2048;
c@42:         config.nceps = m_featureColumnSize - 1;
c@42:         config.want_c0 = true;
c@42:         m_mfcc = new MFCC(config);
c@42:         m_fftSize = m_mfcc->getfftlength();
c@42: 
c@43:         std::cerr << "MFCC FS = " << config.FS << ", FFT size = " << m_fftSize<< std::endl;
c@43: 
c@42:     } else if (m_type == TypeChroma) {
c@42: 
c@42:         m_featureColumnSize = 12;
c@42: 
c@42:         ChromaConfig config;
c@42:         config.FS = lrintf(m_inputSampleRate) / decimationFactor;
c@42:         config.min = Pitch::getFrequencyForPitch(24, 0, 440);
c@42:         config.max = Pitch::getFrequencyForPitch(96, 0, 440);
c@42:         config.BPO = 12;
c@42:         config.CQThresh = 0.0054;
c@42:         config.isNormalised = true;
c@42:         m_chromagram = new Chromagram(config);
c@42:         m_fftSize = m_chromagram->getFrameSize();
c@42: 
c@42:         std::cerr << "min = "<< config.min << ", max = " << config.max << std::endl;
c@42: 
c@42:     } else {
c@42: 
c@42:         std::cerr << "SimilarityPlugin::initialise: internal error: unknown type " << m_type << std::endl;
c@42:         return false;
c@42:     }
c@41:     
c@41:     for (int i = 0; i < m_channels; ++i) {
c@42:         m_values.push_back(FeatureMatrix());
c@41:     }
c@41: 
c@41:     return true;
c@41: }
c@41: 
c@41: void
c@41: SimilarityPlugin::reset()
c@41: {
c@41:     //!!!
c@41: }
c@41: 
c@41: int
c@41: SimilarityPlugin::getDecimationFactor() const
c@41: {
c@41:     int rate = lrintf(m_inputSampleRate);
c@41:     int internalRate = 22050;
c@41:     int decimationFactor = rate / internalRate;
c@41:     if (decimationFactor < 1) decimationFactor = 1;
c@41: 
c@41:     // must be a power of two
c@41:     while (decimationFactor & (decimationFactor - 1)) ++decimationFactor;
c@41: 
c@41:     return decimationFactor;
c@41: }
c@41: 
c@41: size_t
c@41: SimilarityPlugin::getPreferredStepSize() const
c@41: {
c@42:     if (m_blockSize == 0) calculateBlockSize();
c@43:     if (m_type == TypeChroma) {
c@43:         return m_blockSize/2;
c@43:     } else {
c@43:         // for compatibility with old-skool Soundbite, which doesn't
c@43:         // overlap blocks on input
c@43:         return m_blockSize;
c@43:     }
c@41: }
c@41: 
c@41: size_t
c@41: SimilarityPlugin::getPreferredBlockSize() const
c@41: {
c@42:     if (m_blockSize == 0) calculateBlockSize();
c@42:     return m_blockSize;
c@42: }
c@42: 
c@42: void
c@42: SimilarityPlugin::calculateBlockSize() const
c@42: {
c@42:     if (m_blockSize != 0) return;
c@42:     int decimationFactor = getDecimationFactor();
c@42:     if (m_type == TypeChroma) {
c@42:         ChromaConfig config;
c@42:         config.FS = lrintf(m_inputSampleRate) / decimationFactor;
c@42:         config.min = Pitch::getFrequencyForPitch(24, 0, 440);
c@42:         config.max = Pitch::getFrequencyForPitch(96, 0, 440);
c@42:         config.BPO = 12;
c@42:         config.CQThresh = 0.0054;
c@42:         config.isNormalised = false;
c@42:         Chromagram *c = new Chromagram(config);
c@42:         size_t sz = c->getFrameSize();
c@42:         delete c;
c@42:         m_blockSize = sz * decimationFactor;
c@42:     } else {
c@42:         m_blockSize = 2048 * decimationFactor;
c@42:     }
c@41: }
c@41: 
c@41: SimilarityPlugin::ParameterList SimilarityPlugin::getParameterDescriptors() const
c@41: {
c@41:     ParameterList list;
c@42: 
c@42:     ParameterDescriptor desc;
c@42:     desc.identifier = "featureType";
c@42:     desc.name = "Feature Type";
c@42:     desc.description = "";//!!!
c@42:     desc.unit = "";
c@42:     desc.minValue = 0;
c@42:     desc.maxValue = 1;
c@42:     desc.defaultValue = 0;
c@42:     desc.isQuantized = true;
c@42:     desc.quantizeStep = 1;
c@42:     desc.valueNames.push_back("Timbral (MFCC)");
c@42:     desc.valueNames.push_back("Chromatic (Chroma)");
c@42:     list.push_back(desc);	
c@42: 	
c@41:     return list;
c@41: }
c@41: 
c@41: float
c@41: SimilarityPlugin::getParameter(std::string param) const
c@41: {
c@42:     if (param == "featureType") {
c@42:         if (m_type == TypeMFCC) return 0;
c@42:         else if (m_type == TypeChroma) return 1;
c@42:         else return 0;
c@42:     }
c@42: 
c@41:     std::cerr << "WARNING: SimilarityPlugin::getParameter: unknown parameter \""
c@41:               << param << "\"" << std::endl;
c@41:     return 0.0;
c@41: }
c@41: 
c@41: void
c@41: SimilarityPlugin::setParameter(std::string param, float value)
c@41: {
c@42:     if (param == "featureType") {
c@42:         int v = int(value + 0.1);
c@42:         Type prevType = m_type;
c@42:         if (v == 0) m_type = TypeMFCC;
c@42:         else if (v == 1) m_type = TypeChroma;
c@42:         if (m_type != prevType) m_blockSize = 0;
c@42:         return;
c@42:     }
c@42: 
c@41:     std::cerr << "WARNING: SimilarityPlugin::setParameter: unknown parameter \""
c@41:               << param << "\"" << std::endl;
c@41: }
c@41: 
c@41: SimilarityPlugin::OutputList
c@41: SimilarityPlugin::getOutputDescriptors() const
c@41: {
c@41:     OutputList list;
c@41: 	
c@41:     OutputDescriptor similarity;
c@43:     similarity.identifier = "distancematrix";
c@43:     similarity.name = "Distance Matrix";
c@43:     similarity.description = "Distance matrix for similarity metric.  Smaller = more similar.  Should be symmetrical.";
c@41:     similarity.unit = "";
c@41:     similarity.hasFixedBinCount = true;
c@41:     similarity.binCount = m_channels;
c@41:     similarity.hasKnownExtents = false;
c@41:     similarity.isQuantized = false;
c@41:     similarity.sampleType = OutputDescriptor::FixedSampleRate;
c@41:     similarity.sampleRate = 1;
c@41: 	
c@43:     m_distanceMatrixOutput = list.size();
c@41:     list.push_back(similarity);
c@41: 	
c@43:     OutputDescriptor simvec;
c@43:     simvec.identifier = "distancevector";
c@43:     simvec.name = "Distance from First Channel";
c@43:     simvec.description = "Distance vector for similarity of each channel to the first channel.  Smaller = more similar.";
c@43:     simvec.unit = "";
c@43:     simvec.hasFixedBinCount = true;
c@43:     simvec.binCount = m_channels;
c@43:     simvec.hasKnownExtents = false;
c@43:     simvec.isQuantized = false;
c@43:     simvec.sampleType = OutputDescriptor::FixedSampleRate;
c@43:     simvec.sampleRate = 1;
c@43: 	
c@43:     m_distanceVectorOutput = list.size();
c@43:     list.push_back(simvec);
c@43: 	
c@44:     OutputDescriptor sortvec;
c@44:     sortvec.identifier = "sorteddistancevector";
c@44:     sortvec.name = "Ordered Distances from First Channel";
c@44:     sortvec.description = "Vector of the order of other channels in similarity to the first, followed by distance vector for similarity of each to the first.  Smaller = more similar.";
c@44:     sortvec.unit = "";
c@44:     sortvec.hasFixedBinCount = true;
c@44:     sortvec.binCount = m_channels;
c@44:     sortvec.hasKnownExtents = false;
c@44:     sortvec.isQuantized = false;
c@44:     sortvec.sampleType = OutputDescriptor::FixedSampleRate;
c@44:     sortvec.sampleRate = 1;
c@44: 	
c@44:     m_sortedVectorOutput = list.size();
c@44:     list.push_back(sortvec);
c@44: 	
c@41:     OutputDescriptor means;
c@41:     means.identifier = "means";
c@42:     means.name = "Feature Means";
c@43:     means.description = "Means of the feature bins.  Feature time (sec) corresponds to input channel.  Number of bins depends on selected feature type.";
c@41:     means.unit = "";
c@41:     means.hasFixedBinCount = true;
c@43:     means.binCount = m_featureColumnSize;
c@41:     means.hasKnownExtents = false;
c@41:     means.isQuantized = false;
c@43:     means.sampleType = OutputDescriptor::FixedSampleRate;
c@43:     means.sampleRate = 1;
c@41: 	
c@43:     m_meansOutput = list.size();
c@41:     list.push_back(means);
c@41:     
c@41:     OutputDescriptor variances;
c@41:     variances.identifier = "variances";
c@42:     variances.name = "Feature Variances";
c@43:     variances.description = "Variances of the feature bins.  Feature time (sec) corresponds to input channel.  Number of bins depends on selected feature type.";
c@41:     variances.unit = "";
c@41:     variances.hasFixedBinCount = true;
c@43:     variances.binCount = m_featureColumnSize;
c@41:     variances.hasKnownExtents = false;
c@41:     variances.isQuantized = false;
c@43:     variances.sampleType = OutputDescriptor::FixedSampleRate;
c@43:     variances.sampleRate = 1;
c@41: 	
c@43:     m_variancesOutput = list.size();
c@41:     list.push_back(variances);
c@41:     
c@41:     return list;
c@41: }
c@41: 
c@41: SimilarityPlugin::FeatureSet
c@41: SimilarityPlugin::process(const float *const *inputBuffers, Vamp::RealTime /* timestamp */)
c@41: {
c@41:     double *dblbuf = new double[m_blockSize];
c@41:     double *decbuf = dblbuf;
c@42:     if (m_decimator) decbuf = new double[m_fftSize];
c@42: 
c@42:     double *raw = 0;
c@42:     bool ownRaw = false;
c@42: 
c@42:     if (m_type == TypeMFCC) {
c@42:         raw = new double[m_featureColumnSize];
c@42:         ownRaw = true;
c@42:     }
c@41: 
c@43:     float threshold = 1e-10;
c@43: 
c@41:     for (size_t c = 0; c < m_channels; ++c) {
c@41: 
c@43:         bool empty = true;
c@43: 
c@41:         for (int i = 0; i < m_blockSize; ++i) {
c@43:             float val = inputBuffers[c][i];
c@43:             if (fabs(val) > threshold) empty = false;
c@43:             dblbuf[i] = val;
c@41:         }
c@41: 
c@43:         if (empty) continue;
c@44:         m_lastNonEmptyFrame[c] = m_frameNo;
c@43: 
c@41:         if (m_decimator) {
c@41:             m_decimator->process(dblbuf, decbuf);
c@41:         }
c@42: 
c@42:         if (m_type == TypeMFCC) {
c@42:             m_mfcc->process(m_fftSize, decbuf, raw);
c@42:         } else if (m_type == TypeChroma) {
c@42:             raw = m_chromagram->process(decbuf);
c@42:         }                
c@41:         
c@42:         FeatureColumn mf(m_featureColumnSize);
c@44: //        std::cout << m_frameNo << ":" << c << ": ";
c@44:         for (int i = 0; i < m_featureColumnSize; ++i) {
c@44:             mf[i] = raw[i];
c@44: //            std::cout << raw[i] << " ";
c@44:         }
c@44: //        std::cout << std::endl;
c@41: 
c@42:         m_values[c].push_back(mf);
c@41:     }
c@41: 
c@41:     if (m_decimator) delete[] decbuf;
c@41:     delete[] dblbuf;
c@42: 
c@42:     if (ownRaw) delete[] raw;
c@41: 	
c@44:     ++m_frameNo;
c@44: 
c@41:     return FeatureSet();
c@41: }
c@41: 
c@41: SimilarityPlugin::FeatureSet
c@41: SimilarityPlugin::getRemainingFeatures()
c@41: {
c@42:     std::vector<FeatureColumn> m(m_channels);
c@42:     std::vector<FeatureColumn> v(m_channels);
c@41:     
c@41:     for (int i = 0; i < m_channels; ++i) {
c@41: 
c@42:         FeatureColumn mean(m_featureColumnSize), variance(m_featureColumnSize);
c@41: 
c@42:         for (int j = 0; j < m_featureColumnSize; ++j) {
c@41: 
c@43:             mean[j] = 0.0;
c@43:             variance[j] = 0.0;
c@41:             int count;
c@41: 
c@44:             // We want to take values up to, but not including, the
c@44:             // last non-empty frame (which may be partial)
c@43: 
c@44:             int sz = m_lastNonEmptyFrame[i];
c@44:             if (sz < 0) sz = 0;
c@43: 
c@43: //            std::cout << "\nBin " << j << ":" << std::endl;
c@42: 
c@41:             count = 0;
c@43:             for (int k = 0; k < sz; ++k) {
c@42:                 double val = m_values[i][k][j];
c@42: //                std::cout << val << " ";
c@41:                 if (isnan(val) || isinf(val)) continue;
c@41:                 mean[j] += val;
c@41:                 ++count;
c@41:             }
c@41:             if (count > 0) mean[j] /= count;
c@43: //            std::cout << "\n" << count << " non-NaN non-inf values, so mean = " << mean[j] << std::endl;
c@41: 
c@41:             count = 0;
c@43:             for (int k = 0; k < sz; ++k) {
c@42:                 double val = ((m_values[i][k][j] - mean[j]) *
c@42:                               (m_values[i][k][j] - mean[j]));
c@41:                 if (isnan(val) || isinf(val)) continue;
c@41:                 variance[j] += val;
c@41:                 ++count;
c@41:             }
c@41:             if (count > 0) variance[j] /= count;
c@43: //            std::cout << "... and variance = " << variance[j] << std::endl;
c@41:         }
c@41: 
c@41:         m[i] = mean;
c@41:         v[i] = variance;
c@41:     }
c@41: 
c@42:     // we want to return a matrix of the distances between channels,
c@41:     // but Vamp doesn't have a matrix return type so we actually
c@41:     // return a series of vectors
c@41: 
c@41:     std::vector<std::vector<double> > distances;
c@41: 
c@42:     // "Despite the fact that MFCCs extracted from music are clearly
c@42:     // not Gaussian, [14] showed, somewhat surprisingly, that a
c@42:     // similarity function comparing single Gaussians modelling MFCCs
c@42:     // for each track can perform as well as mixture models.  A great
c@42:     // advantage of using single Gaussians is that a simple closed
c@42:     // form exists for the KL divergence." -- Mark Levy, "Lightweight
c@42:     // measures for timbral similarity of musical audio"
c@42:     // (http://www.elec.qmul.ac.uk/easaier/papers/mlevytimbralsimilarity.pdf)
c@42:     //
c@42:     // This code calculates a symmetrised distance metric based on the
c@42:     // KL divergence of Gaussian models of the MFCC values.
c@42: 
c@41:     for (int i = 0; i < m_channels; ++i) {
c@41:         distances.push_back(std::vector<double>());
c@41:         for (int j = 0; j < m_channels; ++j) {
c@42:             double d = -2.0 * m_featureColumnSize;
c@42:             for (int k = 0; k < m_featureColumnSize; ++k) {
c@42:                 // m[i][k] is the mean of feature bin k for channel i
c@42:                 // v[i][k] is the variance of feature bin k for channel i
c@41:                 d += v[i][k] / v[j][k] + v[j][k] / v[i][k];
c@41:                 d += (m[i][k] - m[j][k])
c@41:                     * (1.0 / v[i][k] + 1.0 / v[j][k])
c@41:                     * (m[i][k] - m[j][k]);
c@41:             }
c@41:             d /= 2.0;
c@41:             distances[i].push_back(d);
c@41:         }
c@41:     }
c@41: 
c@44:     // We give all features a timestamp, otherwise hosts will tend to
c@44:     // stamp them at the end of the file, which is annoying
c@44: 
c@41:     FeatureSet returnFeatures;
c@41: 
c@44:     Feature feature;
c@44:     feature.hasTimestamp = true;
c@44: 
c@43:     Feature distanceVectorFeature;
c@43:     distanceVectorFeature.label = "Distance from first channel";
c@44:     distanceVectorFeature.hasTimestamp = true;
c@44:     distanceVectorFeature.timestamp = Vamp::RealTime::zeroTime;
c@44: 
c@44:     std::map<double, int> sorted;
c@44: 
c@44:     char labelBuffer[100];
c@43: 
c@41:     for (int i = 0; i < m_channels; ++i) {
c@41: 
c@41:         feature.timestamp = Vamp::RealTime(i, 0);
c@41: 
c@44:         sprintf(labelBuffer, "Means for channel %d", i+1);
c@44:         feature.label = labelBuffer;
c@44: 
c@41:         feature.values.clear();
c@42:         for (int k = 0; k < m_featureColumnSize; ++k) {
c@41:             feature.values.push_back(m[i][k]);
c@41:         }
c@41: 
c@43:         returnFeatures[m_meansOutput].push_back(feature);
c@41: 
c@44:         sprintf(labelBuffer, "Variances for channel %d", i+1);
c@44:         feature.label = labelBuffer;
c@44: 
c@41:         feature.values.clear();
c@42:         for (int k = 0; k < m_featureColumnSize; ++k) {
c@41:             feature.values.push_back(v[i][k]);
c@41:         }
c@41: 
c@43:         returnFeatures[m_variancesOutput].push_back(feature);
c@41: 
c@41:         feature.values.clear();
c@41:         for (int j = 0; j < m_channels; ++j) {
c@41:             feature.values.push_back(distances[i][j]);
c@41:         }
c@43: 
c@44:         sprintf(labelBuffer, "Distances from channel %d", i+1);
c@44:         feature.label = labelBuffer;
c@41: 		
c@43:         returnFeatures[m_distanceMatrixOutput].push_back(feature);
c@43: 
c@43:         distanceVectorFeature.values.push_back(distances[0][i]);
c@44: 
c@44:         sorted[distances[0][i]] = i;
c@41:     }
c@41: 
c@43:     returnFeatures[m_distanceVectorOutput].push_back(distanceVectorFeature);
c@43: 
c@44:     feature.label = "Order of channels by similarity to first channel";
c@44:     feature.values.clear();
c@44:     feature.timestamp = Vamp::RealTime(0, 0);
c@44: 
c@44:     for (std::map<double, int>::iterator i = sorted.begin();
c@44:          i != sorted.end(); ++i) {
c@44:         feature.values.push_back(i->second);
c@44:     }
c@44: 
c@44:     returnFeatures[m_sortedVectorOutput].push_back(feature);
c@44: 
c@44:     feature.label = "Ordered distances of channels from first channel";
c@44:     feature.values.clear();
c@44:     feature.timestamp = Vamp::RealTime(1, 0);
c@44: 
c@44:     for (std::map<double, int>::iterator i = sorted.begin();
c@44:          i != sorted.end(); ++i) {
c@44:         feature.values.push_back(i->first);
c@44:     }
c@44: 
c@44:     returnFeatures[m_sortedVectorOutput].push_back(feature);
c@44: 
c@41:     return returnFeatures;
c@41: }