Chris@49: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
Chris@0: 
Chris@0: /*
Chris@52:     Sonic Visualiser
Chris@52:     An audio file viewer and annotation editor.
Chris@52:     Centre for Digital Music, Queen Mary, University of London.
Chris@52:     This file copyright 2006 Chris Cannam.
Chris@0:     
Chris@52:     This program is free software; you can redistribute it and/or
Chris@52:     modify it under the terms of the GNU General Public License as
Chris@52:     published by the Free Software Foundation; either version 2 of the
Chris@52:     License, or (at your option) any later version.  See the file
Chris@52:     COPYING included with this distribution for more information.
Chris@0: */
Chris@0: 
Chris@0: #include "FeatureExtractionPluginTransform.h"
Chris@0: 
Chris@0: #include "plugin/FeatureExtractionPluginFactory.h"
Chris@66: #include "plugin/PluginXml.h"
Chris@66: #include "vamp-sdk/Plugin.h"
Chris@0: 
Chris@0: #include "base/Model.h"
Chris@67: #include "base/Window.h"
Chris@0: #include "model/SparseOneDimensionalModel.h"
Chris@0: #include "model/SparseTimeValueModel.h"
Chris@0: #include "model/DenseThreeDimensionalModel.h"
Chris@0: #include "model/DenseTimeValueModel.h"
Chris@115: #include "model/NoteModel.h"
Chris@132: #include "fileio/FFTFuzzyAdapter.h"
Chris@128: 
Chris@67: #include <fftw3.h>
Chris@67: 
Chris@0: #include <iostream>
Chris@0: 
Chris@0: FeatureExtractionPluginTransform::FeatureExtractionPluginTransform(Model *inputModel,
Chris@0: 								   QString pluginId,
Chris@64:                                                                    int channel,
Chris@56:                                                                    QString configurationXml,
Chris@0: 								   QString outputName) :
Chris@0:     Transform(inputModel),
Chris@0:     m_plugin(0),
Chris@64:     m_channel(channel),
Chris@68:     m_stepSize(0),
Chris@68:     m_blockSize(0),
Chris@0:     m_descriptor(0),
Chris@0:     m_outputFeatureNo(0)
Chris@0: {
Chris@117: //    std::cerr << "FeatureExtractionPluginTransform::FeatureExtractionPluginTransform: plugin " << pluginId.toStdString() << ", outputName " << outputName.toStdString() << std::endl;
Chris@0: 
Chris@0:     FeatureExtractionPluginFactory *factory =
Chris@0: 	FeatureExtractionPluginFactory::instanceFor(pluginId);
Chris@0: 
Chris@0:     if (!factory) {
Chris@0: 	std::cerr << "FeatureExtractionPluginTransform: No factory available for plugin id \""
Chris@0: 		  << pluginId.toStdString() << "\"" << std::endl;
Chris@0: 	return;
Chris@0:     }
Chris@0: 
Chris@0:     m_plugin = factory->instantiatePlugin(pluginId, m_input->getSampleRate());
Chris@0: 
Chris@0:     if (!m_plugin) {
Chris@0: 	std::cerr << "FeatureExtractionPluginTransform: Failed to instantiate plugin \""
Chris@0: 		  << pluginId.toStdString() << "\"" << std::endl;
Chris@0: 	return;
Chris@0:     }
Chris@0: 
Chris@56:     if (configurationXml != "") {
Chris@66:         PluginXml(m_plugin).setParametersFromXml(configurationXml);
Chris@56:     }
Chris@56: 
Chris@68:     m_blockSize = m_plugin->getPreferredBlockSize();
Chris@68:     m_stepSize = m_plugin->getPreferredStepSize();
Chris@68: 
Chris@68:     if (m_blockSize == 0) m_blockSize = 1024; //!!! todo: ask user
Chris@68:     if (m_stepSize == 0) m_stepSize = m_blockSize; //!!! likewise
Chris@68: 
Chris@114:     DenseTimeValueModel *input = getInput();
Chris@114:     if (!input) return;
Chris@114: 
Chris@114:     size_t channelCount = input->getChannelCount();
Chris@114:     if (m_plugin->getMaxChannelCount() < channelCount) {
Chris@114: 	channelCount = 1;
Chris@114:     }
Chris@114:     if (m_plugin->getMinChannelCount() > channelCount) {
Chris@114: 	std::cerr << "FeatureExtractionPluginTransform:: "
Chris@114: 		  << "Can't provide enough channels to plugin (plugin min "
Chris@114: 		  << m_plugin->getMinChannelCount() << ", max "
Chris@114: 		  << m_plugin->getMaxChannelCount() << ", input model has "
Chris@114: 		  << input->getChannelCount() << ")" << std::endl;
Chris@114: 	return;
Chris@114:     }
Chris@114: 
Chris@114:     if (!m_plugin->initialise(channelCount, m_stepSize, m_blockSize)) {
Chris@114:         std::cerr << "FeatureExtractionPluginTransform: Plugin "
Chris@114:                   << m_plugin->getName() << " failed to initialise!" << std::endl;
Chris@114:         return;
Chris@114:     }
Chris@114: 
Chris@74:     Vamp::Plugin::OutputList outputs = m_plugin->getOutputDescriptors();
Chris@0: 
Chris@0:     if (outputs.empty()) {
Chris@0: 	std::cerr << "FeatureExtractionPluginTransform: Plugin \""
Chris@0: 		  << pluginId.toStdString() << "\" has no outputs" << std::endl;
Chris@0: 	return;
Chris@0:     }
Chris@0:     
Chris@0:     for (size_t i = 0; i < outputs.size(); ++i) {
Chris@0: 	if (outputName == "" || outputs[i].name == outputName.toStdString()) {
Chris@0: 	    m_outputFeatureNo = i;
Chris@66: 	    m_descriptor = new Vamp::Plugin::OutputDescriptor
Chris@0: 		(outputs[i]);
Chris@0: 	    break;
Chris@0: 	}
Chris@0:     }
Chris@0: 
Chris@0:     if (!m_descriptor) {
Chris@0: 	std::cerr << "FeatureExtractionPluginTransform: Plugin \""
Chris@0: 		  << pluginId.toStdString() << "\" has no output named \""
Chris@0: 		  << outputName.toStdString() << "\"" << std::endl;
Chris@0: 	return;
Chris@0:     }
Chris@0: 
Chris@117: //    std::cerr << "FeatureExtractionPluginTransform: output sample type "
Chris@117: //	      << m_descriptor->sampleType << std::endl;
Chris@0: 
Chris@70:     int binCount = 1;
Chris@0:     float minValue = 0.0, maxValue = 0.0;
Chris@0:     
Chris@70:     if (m_descriptor->hasFixedBinCount) {
Chris@70: 	binCount = m_descriptor->binCount;
Chris@0:     }
Chris@0: 
Chris@117: //    std::cerr << "FeatureExtractionPluginTransform: output bin count "
Chris@117: //	      << binCount << std::endl;
Chris@114: 
Chris@70:     if (binCount > 0 && m_descriptor->hasKnownExtents) {
Chris@0: 	minValue = m_descriptor->minValue;
Chris@0: 	maxValue = m_descriptor->maxValue;
Chris@0:     }
Chris@0: 
Chris@0:     size_t modelRate = m_input->getSampleRate();
Chris@0:     size_t modelResolution = 1;
Chris@0:     
Chris@0:     switch (m_descriptor->sampleType) {
Chris@0: 
Chris@66:     case Vamp::Plugin::OutputDescriptor::VariableSampleRate:
Chris@0: 	if (m_descriptor->sampleRate != 0.0) {
Chris@0: 	    modelResolution = size_t(modelRate / m_descriptor->sampleRate + 0.001);
Chris@0: 	}
Chris@0: 	break;
Chris@0: 
Chris@66:     case Vamp::Plugin::OutputDescriptor::OneSamplePerStep:
Chris@72: 	modelResolution = m_stepSize;
Chris@0: 	break;
Chris@0: 
Chris@66:     case Vamp::Plugin::OutputDescriptor::FixedSampleRate:
Chris@73: 	modelRate = size_t(m_descriptor->sampleRate + 0.001);
Chris@0: 	break;
Chris@0:     }
Chris@0: 
Chris@70:     if (binCount == 0) {
Chris@0: 
Chris@20: 	m_output = new SparseOneDimensionalModel(modelRate, modelResolution,
Chris@20: 						 false);
Chris@0: 
Chris@115:     } else if (binCount == 1) {
Chris@0: 
Chris@115:         SparseTimeValueModel *model = new SparseTimeValueModel
Chris@115:             (modelRate, modelResolution, minValue, maxValue, false);
Chris@115:         model->setScaleUnits(outputs[m_outputFeatureNo].unit.c_str());
Chris@115: 
Chris@115:         m_output = model;
Chris@115: 
Chris@115:     } else if (m_descriptor->sampleType ==
Chris@66: 	       Vamp::Plugin::OutputDescriptor::VariableSampleRate) {
Chris@115: 
Chris@115:         // We don't have a sparse 3D model, so interpret this as a
Chris@115:         // note model.  There's nothing to define which values to use
Chris@115:         // as which parameters of the note -- for the moment let's
Chris@115:         // treat the first as pitch, second as duration in frames,
Chris@115:         // third (if present) as velocity. (Our note model doesn't
Chris@115:         // yet store velocity.)
Chris@115:         //!!! todo: ask the user!
Chris@0: 	
Chris@115:         NoteModel *model = new NoteModel
Chris@63:             (modelRate, modelResolution, minValue, maxValue, false);
Chris@63:         model->setScaleUnits(outputs[m_outputFeatureNo].unit.c_str());
Chris@63: 
Chris@63:         m_output = model;
Chris@0: 
Chris@0:     } else {
Chris@0: 	
Chris@0: 	m_output = new DenseThreeDimensionalModel(modelRate, modelResolution,
Chris@70: 						  binCount, false);
Chris@20: 
Chris@70: 	if (!m_descriptor->binNames.empty()) {
Chris@20: 	    std::vector<QString> names;
Chris@70: 	    for (size_t i = 0; i < m_descriptor->binNames.size(); ++i) {
Chris@70: 		names.push_back(m_descriptor->binNames[i].c_str());
Chris@20: 	    }
Chris@20: 	    (dynamic_cast<DenseThreeDimensionalModel *>(m_output))
Chris@20: 		->setBinNames(names);
Chris@20: 	}
Chris@0:     }
Chris@0: }
Chris@0: 
Chris@0: FeatureExtractionPluginTransform::~FeatureExtractionPluginTransform()
Chris@0: {
Chris@0:     delete m_plugin;
Chris@0:     delete m_descriptor;
Chris@0: }
Chris@0: 
Chris@0: DenseTimeValueModel *
Chris@0: FeatureExtractionPluginTransform::getInput()
Chris@0: {
Chris@0:     DenseTimeValueModel *dtvm =
Chris@0: 	dynamic_cast<DenseTimeValueModel *>(getInputModel());
Chris@0:     if (!dtvm) {
Chris@0: 	std::cerr << "FeatureExtractionPluginTransform::getInput: WARNING: Input model is not conformable to DenseTimeValueModel" << std::endl;
Chris@0:     }
Chris@0:     return dtvm;
Chris@0: }
Chris@0: 
Chris@0: void
Chris@0: FeatureExtractionPluginTransform::run()
Chris@0: {
Chris@0:     DenseTimeValueModel *input = getInput();
Chris@0:     if (!input) return;
Chris@0: 
Chris@0:     if (!m_output) return;
Chris@0: 
Chris@114:     size_t sampleRate = m_input->getSampleRate();
Chris@114: 
Chris@0:     size_t channelCount = input->getChannelCount();
Chris@0:     if (m_plugin->getMaxChannelCount() < channelCount) {
Chris@0: 	channelCount = 1;
Chris@0:     }
Chris@0: 
Chris@0:     float **buffers = new float*[channelCount];
Chris@0:     for (size_t ch = 0; ch < channelCount; ++ch) {
Chris@68: 	buffers[ch] = new float[m_blockSize];
Chris@0:     }
Chris@0: 
Chris@128:     bool frequencyDomain = (m_plugin->getInputDomain() ==
Chris@128:                             Vamp::Plugin::FrequencyDomain);
Chris@132:     std::vector<FFTFuzzyAdapter *> fftAdapters;
Chris@131: 
Chris@128:     if (frequencyDomain) {
Chris@128:         for (size_t ch = 0; ch < channelCount; ++ch) {
Chris@132:             fftAdapters.push_back(new FFTFuzzyAdapter
Chris@132:                                   (getInput(),
Chris@132:                                    channelCount == 1 ? m_channel : ch,
Chris@132:                                    HanningWindow,
Chris@132:                                    m_blockSize,
Chris@132:                                    m_stepSize,
Chris@132:                                    m_blockSize,
Chris@132:                                    false));
Chris@128:         }
Chris@128:     }
Chris@67: 
Chris@70:     long startFrame = m_input->getStartFrame();
Chris@70:     long   endFrame = m_input->getEndFrame();
Chris@70:     long blockFrame = startFrame;
Chris@0: 
Chris@70:     long prevCompletion = 0;
Chris@0: 
Chris@70:     while (1) {
Chris@70: 
Chris@128:         if (frequencyDomain) {
Chris@73:             if (blockFrame - int(m_blockSize)/2 > endFrame) break;
Chris@70:         } else {
Chris@70:             if (blockFrame >= endFrame) break;
Chris@70:         }
Chris@0: 
Chris@0: //	std::cerr << "FeatureExtractionPluginTransform::run: blockFrame "
Chris@0: //		  << blockFrame << std::endl;
Chris@0: 
Chris@70: 	long completion =
Chris@68: 	    (((blockFrame - startFrame) / m_stepSize) * 99) /
Chris@68: 	    (   (endFrame - startFrame) / m_stepSize);
Chris@0: 
Chris@0: 	// channelCount is either m_input->channelCount or 1
Chris@0: 
Chris@73:         for (size_t ch = 0; ch < channelCount; ++ch) {
Chris@128:             if (frequencyDomain) {
Chris@128:                 int column = (blockFrame - startFrame) / m_stepSize;
Chris@128:                 for (size_t i = 0; i < m_blockSize/2; ++i) {
Chris@132:                     fftAdapters[ch]->getValuesAt
Chris@128:                         (column, i, buffers[ch][i*2], buffers[ch][i*2+1]);
Chris@128:                 }
Chris@133:                 //!!!
Chris@133:                 float sum = 0.0;
Chris@133:                 for (size_t i = 0; i < m_blockSize/2; ++i) {
Chris@133:                     sum += buffers[ch][i*2];
Chris@133:                 }
Chris@133:                 if (fabs(sum) < 0.0001) {
Chris@133:                     std::cerr << "WARNING: small sum for column " << column << " (sum is " << sum << ")" << std::endl;
Chris@133:                 }
Chris@70:             } else {
Chris@70:                 getFrames(ch, channelCount, 
Chris@128:                           blockFrame, m_blockSize, buffers[ch]);
Chris@70:             }                
Chris@70:         }
Chris@132: 
Chris@66: 	Vamp::Plugin::FeatureSet features = m_plugin->process
Chris@66: 	    (buffers, Vamp::RealTime::frame2RealTime(blockFrame, sampleRate));
Chris@0: 
Chris@0: 	for (size_t fi = 0; fi < features[m_outputFeatureNo].size(); ++fi) {
Chris@66: 	    Vamp::Plugin::Feature feature =
Chris@0: 		features[m_outputFeatureNo][fi];
Chris@0: 	    addFeature(blockFrame, feature);
Chris@0: 	}
Chris@0: 
Chris@0: 	if (blockFrame == startFrame || completion > prevCompletion) {
Chris@0: 	    setCompletion(completion);
Chris@0: 	    prevCompletion = completion;
Chris@0: 	}
Chris@0: 
Chris@68: 	blockFrame += m_stepSize;
Chris@0:     }
Chris@132: 
Chris@66:     Vamp::Plugin::FeatureSet features = m_plugin->getRemainingFeatures();
Chris@0: 
Chris@0:     for (size_t fi = 0; fi < features[m_outputFeatureNo].size(); ++fi) {
Chris@66: 	Vamp::Plugin::Feature feature =
Chris@0: 	    features[m_outputFeatureNo][fi];
Chris@0: 	addFeature(blockFrame, feature);
Chris@0:     }
Chris@0: 
Chris@132:     if (frequencyDomain) {
Chris@132:         for (size_t ch = 0; ch < channelCount; ++ch) {
Chris@132:             delete fftAdapters[ch];
Chris@132:         }
Chris@132:     }
Chris@132: 
Chris@0:     setCompletion(100);
Chris@0: }
Chris@0: 
Chris@70: void
Chris@70: FeatureExtractionPluginTransform::getFrames(int channel, int channelCount,
Chris@70:                                             long startFrame, long size,
Chris@70:                                             float *buffer)
Chris@70: {
Chris@70:     long offset = 0;
Chris@70: 
Chris@70:     if (startFrame < 0) {
Chris@70:         for (int i = 0; i < size && startFrame + i < 0; ++i) {
Chris@70:             buffer[i] = 0.0f;
Chris@70:         }
Chris@70:         offset = -startFrame;
Chris@70:         size -= offset;
Chris@70:         if (size <= 0) return;
Chris@70:         startFrame = 0;
Chris@70:     }
Chris@70: 
Chris@73:     long got = getInput()->getValues
Chris@70:         ((channelCount == 1 ? m_channel : channel),
Chris@70:          startFrame, startFrame + size, buffer + offset);
Chris@70: 
Chris@70:     while (got < size) {
Chris@70:         buffer[offset + got] = 0.0;
Chris@70:         ++got;
Chris@70:     }
Chris@74: 
Chris@74:     if (m_channel == -1 && channelCount == 1 &&
Chris@74:         getInput()->getChannelCount() > 1) {
Chris@74:         // use mean instead of sum, as plugin input
Chris@74:         int cc = getInput()->getChannelCount();
Chris@74:         for (long i = 0; i < size; ++i) {
Chris@74:             buffer[i] /= cc;
Chris@74:         }
Chris@74:     }
Chris@70: }
Chris@0: 
Chris@0: void
Chris@0: FeatureExtractionPluginTransform::addFeature(size_t blockFrame,
Chris@66: 					     const Vamp::Plugin::Feature &feature)
Chris@0: {
Chris@0:     size_t inputRate = m_input->getSampleRate();
Chris@0: 
Chris@0: //    std::cerr << "FeatureExtractionPluginTransform::addFeature("
Chris@0: //	      << blockFrame << ")" << std::endl;
Chris@0: 
Chris@70:     int binCount = 1;
Chris@70:     if (m_descriptor->hasFixedBinCount) {
Chris@70: 	binCount = m_descriptor->binCount;
Chris@0:     }
Chris@0: 
Chris@0:     size_t frame = blockFrame;
Chris@0: 
Chris@0:     if (m_descriptor->sampleType ==
Chris@66: 	Vamp::Plugin::OutputDescriptor::VariableSampleRate) {
Chris@0: 
Chris@0: 	if (!feature.hasTimestamp) {
Chris@0: 	    std::cerr
Chris@0: 		<< "WARNING: FeatureExtractionPluginTransform::addFeature: "
Chris@0: 		<< "Feature has variable sample rate but no timestamp!"
Chris@0: 		<< std::endl;
Chris@0: 	    return;
Chris@0: 	} else {
Chris@66: 	    frame = Vamp::RealTime::realTime2Frame(feature.timestamp, inputRate);
Chris@0: 	}
Chris@0: 
Chris@0:     } else if (m_descriptor->sampleType ==
Chris@66: 	       Vamp::Plugin::OutputDescriptor::FixedSampleRate) {
Chris@0: 
Chris@0: 	if (feature.hasTimestamp) {
Chris@0: 	    //!!! warning: sampleRate may be non-integral
Chris@66: 	    frame = Vamp::RealTime::realTime2Frame(feature.timestamp,
Chris@66:                                                    m_descriptor->sampleRate);
Chris@0: 	} else {
Chris@0: 	    frame = m_output->getEndFrame() + 1;
Chris@0: 	}
Chris@0:     }
Chris@0: 	
Chris@70:     if (binCount == 0) {
Chris@0: 
Chris@0: 	SparseOneDimensionalModel *model = getOutput<SparseOneDimensionalModel>();
Chris@0: 	if (!model) return;
Chris@0: 	model->addPoint(SparseOneDimensionalModel::Point(frame, feature.label.c_str()));
Chris@0: 	
Chris@115:     } else if (binCount == 1) {
Chris@0: 
Chris@0: 	float value = 0.0;
Chris@0: 	if (feature.values.size() > 0) value = feature.values[0];
Chris@0: 
Chris@0: 	SparseTimeValueModel *model = getOutput<SparseTimeValueModel>();
Chris@0: 	if (!model) return;
Chris@0: 	model->addPoint(SparseTimeValueModel::Point(frame, value, feature.label.c_str()));
Chris@115: 
Chris@115:     } else if (m_descriptor->sampleType == 
Chris@115: 	       Vamp::Plugin::OutputDescriptor::VariableSampleRate) {
Chris@115: 
Chris@115:         float pitch = 0.0;
Chris@115:         if (feature.values.size() > 0) pitch = feature.values[0];
Chris@115: 
Chris@115:         float duration = 1;
Chris@115:         if (feature.values.size() > 1) duration = feature.values[1];
Chris@115:         
Chris@115:         float velocity = 100;
Chris@115:         if (feature.values.size() > 2) velocity = feature.values[2];
Chris@115: 
Chris@115:         NoteModel *model = getOutput<NoteModel>();
Chris@115:         if (!model) return;
Chris@115: 
Chris@115:         model->addPoint(NoteModel::Point(frame, pitch, duration, feature.label.c_str()));
Chris@0: 	
Chris@0:     } else {
Chris@0: 	
Chris@0: 	DenseThreeDimensionalModel::BinValueSet values = feature.values;
Chris@0: 	
Chris@0: 	DenseThreeDimensionalModel *model = getOutput<DenseThreeDimensionalModel>();
Chris@0: 	if (!model) return;
Chris@0: 
Chris@0: 	model->setBinValues(frame, values);
Chris@0:     }
Chris@0: }
Chris@0: 
Chris@0: void
Chris@0: FeatureExtractionPluginTransform::setCompletion(int completion)
Chris@0: {
Chris@70:     int binCount = 1;
Chris@70:     if (m_descriptor->hasFixedBinCount) {
Chris@70: 	binCount = m_descriptor->binCount;
Chris@0:     }
Chris@0: 
Chris@70:     if (binCount == 0) {
Chris@0: 
Chris@0: 	SparseOneDimensionalModel *model = getOutput<SparseOneDimensionalModel>();
Chris@0: 	if (!model) return;
Chris@0: 	model->setCompletion(completion);
Chris@0: 
Chris@115:     } else if (binCount == 1) {
Chris@115: 
Chris@115: 	SparseTimeValueModel *model = getOutput<SparseTimeValueModel>();
Chris@115: 	if (!model) return;
Chris@115: 	model->setCompletion(completion);
Chris@115: 
Chris@115:     } else if (m_descriptor->sampleType ==
Chris@66: 	       Vamp::Plugin::OutputDescriptor::VariableSampleRate) {
Chris@0: 
Chris@115: 	NoteModel *model = getOutput<NoteModel>();
Chris@0: 	if (!model) return;
Chris@0: 	model->setCompletion(completion);
Chris@0: 
Chris@0:     } else {
Chris@0: 
Chris@19: 	DenseThreeDimensionalModel *model = getOutput<DenseThreeDimensionalModel>();
Chris@19: 	if (!model) return;
Chris@19: 	model->setCompletion(completion);
Chris@0:     }
Chris@0: }
Chris@0: