annotate transform/FeatureExtractionPluginTransform.cpp @ 29:660a973c157a

* Handle plugins with weird block size requirements properly
author Chris Cannam
date Wed, 20 Sep 2006 15:47:16 +0000
parents 61259228d029
children 544ab25d2372
rev   line source
Chris@0 1 /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*- vi:set ts=8 sts=4 sw=4: */
Chris@0 2
Chris@0 3 /*
Chris@0 4 Sonic Visualiser
Chris@0 5 An audio file viewer and annotation editor.
Chris@0 6 Centre for Digital Music, Queen Mary, University of London.
Chris@0 7 This file copyright 2006 Chris Cannam.
Chris@0 8
Chris@0 9 This program is free software; you can redistribute it and/or
Chris@0 10 modify it under the terms of the GNU General Public License as
Chris@0 11 published by the Free Software Foundation; either version 2 of the
Chris@0 12 License, or (at your option) any later version. See the file
Chris@0 13 COPYING included with this distribution for more information.
Chris@0 14 */
Chris@0 15
Chris@0 16 #include "FeatureExtractionPluginTransform.h"
Chris@0 17
Chris@0 18 #include "plugin/FeatureExtractionPluginFactory.h"
Chris@0 19 #include "plugin/PluginXml.h"
Chris@0 20 #include "vamp-sdk/Plugin.h"
Chris@0 21
Chris@1 22 #include "data/model/Model.h"
Chris@0 23 #include "base/Window.h"
Chris@1 24 #include "data/model/SparseOneDimensionalModel.h"
Chris@1 25 #include "data/model/SparseTimeValueModel.h"
Chris@3 26 #include "data/model/EditableDenseThreeDimensionalModel.h"
Chris@1 27 #include "data/model/DenseTimeValueModel.h"
Chris@1 28 #include "data/model/NoteModel.h"
Chris@3 29 #include "data/model/FFTModel.h"
Chris@0 30
Chris@0 31 #include <fftw3.h>
Chris@0 32
Chris@0 33 #include <iostream>
Chris@0 34
Chris@0 35 FeatureExtractionPluginTransform::FeatureExtractionPluginTransform(Model *inputModel,
Chris@0 36 QString pluginId,
Chris@27 37 const ExecutionContext &context,
Chris@0 38 QString configurationXml,
Chris@27 39 QString outputName) :
Chris@27 40 PluginTransform(inputModel, context),
Chris@0 41 m_plugin(0),
Chris@0 42 m_descriptor(0),
Chris@0 43 m_outputFeatureNo(0)
Chris@0 44 {
Chris@0 45 // std::cerr << "FeatureExtractionPluginTransform::FeatureExtractionPluginTransform: plugin " << pluginId.toStdString() << ", outputName " << outputName.toStdString() << std::endl;
Chris@0 46
Chris@0 47 FeatureExtractionPluginFactory *factory =
Chris@0 48 FeatureExtractionPluginFactory::instanceFor(pluginId);
Chris@0 49
Chris@0 50 if (!factory) {
Chris@0 51 std::cerr << "FeatureExtractionPluginTransform: No factory available for plugin id \""
Chris@0 52 << pluginId.toStdString() << "\"" << std::endl;
Chris@0 53 return;
Chris@0 54 }
Chris@0 55
Chris@0 56 m_plugin = factory->instantiatePlugin(pluginId, m_input->getSampleRate());
Chris@0 57
Chris@0 58 if (!m_plugin) {
Chris@0 59 std::cerr << "FeatureExtractionPluginTransform: Failed to instantiate plugin \""
Chris@0 60 << pluginId.toStdString() << "\"" << std::endl;
Chris@0 61 return;
Chris@0 62 }
Chris@0 63
Chris@0 64 if (configurationXml != "") {
Chris@0 65 PluginXml(m_plugin).setParametersFromXml(configurationXml);
Chris@0 66 }
Chris@0 67
Chris@0 68 DenseTimeValueModel *input = getInput();
Chris@0 69 if (!input) return;
Chris@0 70
Chris@0 71 size_t channelCount = input->getChannelCount();
Chris@0 72 if (m_plugin->getMaxChannelCount() < channelCount) {
Chris@0 73 channelCount = 1;
Chris@0 74 }
Chris@0 75 if (m_plugin->getMinChannelCount() > channelCount) {
Chris@0 76 std::cerr << "FeatureExtractionPluginTransform:: "
Chris@0 77 << "Can't provide enough channels to plugin (plugin min "
Chris@0 78 << m_plugin->getMinChannelCount() << ", max "
Chris@0 79 << m_plugin->getMaxChannelCount() << ", input model has "
Chris@0 80 << input->getChannelCount() << ")" << std::endl;
Chris@0 81 return;
Chris@0 82 }
Chris@0 83
Chris@27 84 std::cerr << "Initialising feature extraction plugin with channels = "
Chris@27 85 << channelCount << ", step = " << m_context.stepSize
Chris@27 86 << ", block = " << m_context.blockSize << std::endl;
Chris@27 87
Chris@27 88 if (!m_plugin->initialise(channelCount,
Chris@27 89 m_context.stepSize,
Chris@27 90 m_context.blockSize)) {
Chris@0 91 std::cerr << "FeatureExtractionPluginTransform: Plugin "
Chris@0 92 << m_plugin->getName() << " failed to initialise!" << std::endl;
Chris@0 93 return;
Chris@0 94 }
Chris@0 95
Chris@0 96 Vamp::Plugin::OutputList outputs = m_plugin->getOutputDescriptors();
Chris@0 97
Chris@0 98 if (outputs.empty()) {
Chris@0 99 std::cerr << "FeatureExtractionPluginTransform: Plugin \""
Chris@0 100 << pluginId.toStdString() << "\" has no outputs" << std::endl;
Chris@0 101 return;
Chris@0 102 }
Chris@0 103
Chris@0 104 for (size_t i = 0; i < outputs.size(); ++i) {
Chris@0 105 if (outputName == "" || outputs[i].name == outputName.toStdString()) {
Chris@0 106 m_outputFeatureNo = i;
Chris@0 107 m_descriptor = new Vamp::Plugin::OutputDescriptor
Chris@0 108 (outputs[i]);
Chris@0 109 break;
Chris@0 110 }
Chris@0 111 }
Chris@0 112
Chris@0 113 if (!m_descriptor) {
Chris@0 114 std::cerr << "FeatureExtractionPluginTransform: Plugin \""
Chris@0 115 << pluginId.toStdString() << "\" has no output named \""
Chris@0 116 << outputName.toStdString() << "\"" << std::endl;
Chris@0 117 return;
Chris@0 118 }
Chris@0 119
Chris@0 120 // std::cerr << "FeatureExtractionPluginTransform: output sample type "
Chris@0 121 // << m_descriptor->sampleType << std::endl;
Chris@0 122
Chris@0 123 int binCount = 1;
Chris@0 124 float minValue = 0.0, maxValue = 0.0;
Chris@0 125
Chris@0 126 if (m_descriptor->hasFixedBinCount) {
Chris@0 127 binCount = m_descriptor->binCount;
Chris@0 128 }
Chris@0 129
Chris@0 130 // std::cerr << "FeatureExtractionPluginTransform: output bin count "
Chris@0 131 // << binCount << std::endl;
Chris@0 132
Chris@0 133 if (binCount > 0 && m_descriptor->hasKnownExtents) {
Chris@0 134 minValue = m_descriptor->minValue;
Chris@0 135 maxValue = m_descriptor->maxValue;
Chris@0 136 }
Chris@0 137
Chris@0 138 size_t modelRate = m_input->getSampleRate();
Chris@0 139 size_t modelResolution = 1;
Chris@0 140
Chris@0 141 switch (m_descriptor->sampleType) {
Chris@0 142
Chris@0 143 case Vamp::Plugin::OutputDescriptor::VariableSampleRate:
Chris@0 144 if (m_descriptor->sampleRate != 0.0) {
Chris@0 145 modelResolution = size_t(modelRate / m_descriptor->sampleRate + 0.001);
Chris@0 146 }
Chris@0 147 break;
Chris@0 148
Chris@0 149 case Vamp::Plugin::OutputDescriptor::OneSamplePerStep:
Chris@27 150 modelResolution = m_context.stepSize;
Chris@0 151 break;
Chris@0 152
Chris@0 153 case Vamp::Plugin::OutputDescriptor::FixedSampleRate:
Chris@0 154 modelRate = size_t(m_descriptor->sampleRate + 0.001);
Chris@0 155 break;
Chris@0 156 }
Chris@0 157
Chris@0 158 if (binCount == 0) {
Chris@0 159
Chris@0 160 m_output = new SparseOneDimensionalModel(modelRate, modelResolution,
Chris@0 161 false);
Chris@0 162
Chris@0 163 } else if (binCount == 1) {
Chris@0 164
Chris@0 165 SparseTimeValueModel *model = new SparseTimeValueModel
Chris@0 166 (modelRate, modelResolution, minValue, maxValue, false);
Chris@0 167 model->setScaleUnits(outputs[m_outputFeatureNo].unit.c_str());
Chris@0 168
Chris@0 169 m_output = model;
Chris@0 170
Chris@0 171 } else if (m_descriptor->sampleType ==
Chris@0 172 Vamp::Plugin::OutputDescriptor::VariableSampleRate) {
Chris@0 173
Chris@0 174 // We don't have a sparse 3D model, so interpret this as a
Chris@0 175 // note model. There's nothing to define which values to use
Chris@0 176 // as which parameters of the note -- for the moment let's
Chris@0 177 // treat the first as pitch, second as duration in frames,
Chris@0 178 // third (if present) as velocity. (Our note model doesn't
Chris@0 179 // yet store velocity.)
Chris@0 180 //!!! todo: ask the user!
Chris@0 181
Chris@0 182 NoteModel *model = new NoteModel
Chris@0 183 (modelRate, modelResolution, minValue, maxValue, false);
Chris@0 184 model->setScaleUnits(outputs[m_outputFeatureNo].unit.c_str());
Chris@0 185
Chris@0 186 m_output = model;
Chris@0 187
Chris@0 188 } else {
Chris@0 189
Chris@3 190 m_output = new EditableDenseThreeDimensionalModel
Chris@3 191 (modelRate, modelResolution, binCount, false);
Chris@0 192
Chris@0 193 if (!m_descriptor->binNames.empty()) {
Chris@0 194 std::vector<QString> names;
Chris@0 195 for (size_t i = 0; i < m_descriptor->binNames.size(); ++i) {
Chris@0 196 names.push_back(m_descriptor->binNames[i].c_str());
Chris@0 197 }
Chris@3 198 (dynamic_cast<EditableDenseThreeDimensionalModel *>(m_output))
Chris@0 199 ->setBinNames(names);
Chris@0 200 }
Chris@0 201 }
Chris@0 202 }
Chris@0 203
Chris@0 204 FeatureExtractionPluginTransform::~FeatureExtractionPluginTransform()
Chris@0 205 {
Chris@0 206 delete m_plugin;
Chris@0 207 delete m_descriptor;
Chris@0 208 }
Chris@0 209
Chris@0 210 DenseTimeValueModel *
Chris@0 211 FeatureExtractionPluginTransform::getInput()
Chris@0 212 {
Chris@0 213 DenseTimeValueModel *dtvm =
Chris@0 214 dynamic_cast<DenseTimeValueModel *>(getInputModel());
Chris@0 215 if (!dtvm) {
Chris@0 216 std::cerr << "FeatureExtractionPluginTransform::getInput: WARNING: Input model is not conformable to DenseTimeValueModel" << std::endl;
Chris@0 217 }
Chris@0 218 return dtvm;
Chris@0 219 }
Chris@0 220
Chris@0 221 void
Chris@0 222 FeatureExtractionPluginTransform::run()
Chris@0 223 {
Chris@0 224 DenseTimeValueModel *input = getInput();
Chris@0 225 if (!input) return;
Chris@0 226
Chris@0 227 if (!m_output) return;
Chris@0 228
Chris@0 229 size_t sampleRate = m_input->getSampleRate();
Chris@0 230
Chris@0 231 size_t channelCount = input->getChannelCount();
Chris@0 232 if (m_plugin->getMaxChannelCount() < channelCount) {
Chris@0 233 channelCount = 1;
Chris@0 234 }
Chris@0 235
Chris@0 236 float **buffers = new float*[channelCount];
Chris@0 237 for (size_t ch = 0; ch < channelCount; ++ch) {
Chris@27 238 buffers[ch] = new float[m_context.blockSize];
Chris@0 239 }
Chris@0 240
Chris@0 241 bool frequencyDomain = (m_plugin->getInputDomain() ==
Chris@0 242 Vamp::Plugin::FrequencyDomain);
Chris@3 243 std::vector<FFTModel *> fftModels;
Chris@0 244
Chris@0 245 if (frequencyDomain) {
Chris@0 246 for (size_t ch = 0; ch < channelCount; ++ch) {
Chris@3 247 fftModels.push_back(new FFTModel
Chris@0 248 (getInput(),
Chris@27 249 channelCount == 1 ? m_context.channel : ch,
Chris@27 250 m_context.windowType,
Chris@27 251 m_context.blockSize,
Chris@27 252 m_context.stepSize,
Chris@27 253 m_context.blockSize,
Chris@0 254 false));
Chris@0 255 }
Chris@0 256 }
Chris@0 257
Chris@0 258 long startFrame = m_input->getStartFrame();
Chris@0 259 long endFrame = m_input->getEndFrame();
Chris@0 260 long blockFrame = startFrame;
Chris@0 261
Chris@0 262 long prevCompletion = 0;
Chris@0 263
Chris@0 264 while (1) {
Chris@0 265
Chris@0 266 if (frequencyDomain) {
Chris@27 267 if (blockFrame - int(m_context.blockSize)/2 > endFrame) break;
Chris@0 268 } else {
Chris@0 269 if (blockFrame >= endFrame) break;
Chris@0 270 }
Chris@0 271
Chris@29 272 std::cerr << "FeatureExtractionPluginTransform::run: blockFrame "
Chris@29 273 << blockFrame << std::endl;
Chris@0 274
Chris@0 275 long completion =
Chris@27 276 (((blockFrame - startFrame) / m_context.stepSize) * 99) /
Chris@27 277 ( (endFrame - startFrame) / m_context.stepSize);
Chris@0 278
Chris@0 279 // channelCount is either m_input->channelCount or 1
Chris@0 280
Chris@0 281 for (size_t ch = 0; ch < channelCount; ++ch) {
Chris@0 282 if (frequencyDomain) {
Chris@27 283 int column = (blockFrame - startFrame) / m_context.stepSize;
Chris@27 284 for (size_t i = 0; i < m_context.blockSize/2; ++i) {
Chris@3 285 fftModels[ch]->getValuesAt
Chris@0 286 (column, i, buffers[ch][i*2], buffers[ch][i*2+1]);
Chris@0 287 }
Chris@0 288 /*!!!
Chris@0 289 float sum = 0.0;
Chris@27 290 for (size_t i = 0; i < m_context.blockSize/2; ++i) {
Chris@0 291 sum += buffers[ch][i*2];
Chris@0 292 }
Chris@0 293 if (fabs(sum) < 0.0001) {
Chris@0 294 std::cerr << "WARNING: small sum for column " << column << " (sum is " << sum << ")" << std::endl;
Chris@0 295 }
Chris@0 296 */
Chris@0 297 } else {
Chris@0 298 getFrames(ch, channelCount,
Chris@27 299 blockFrame, m_context.blockSize, buffers[ch]);
Chris@0 300 }
Chris@0 301 }
Chris@0 302
Chris@0 303 Vamp::Plugin::FeatureSet features = m_plugin->process
Chris@0 304 (buffers, Vamp::RealTime::frame2RealTime(blockFrame, sampleRate));
Chris@0 305
Chris@0 306 for (size_t fi = 0; fi < features[m_outputFeatureNo].size(); ++fi) {
Chris@0 307 Vamp::Plugin::Feature feature =
Chris@0 308 features[m_outputFeatureNo][fi];
Chris@0 309 addFeature(blockFrame, feature);
Chris@0 310 }
Chris@0 311
Chris@0 312 if (blockFrame == startFrame || completion > prevCompletion) {
Chris@0 313 setCompletion(completion);
Chris@0 314 prevCompletion = completion;
Chris@0 315 }
Chris@0 316
Chris@27 317 blockFrame += m_context.stepSize;
Chris@0 318 }
Chris@0 319
Chris@0 320 Vamp::Plugin::FeatureSet features = m_plugin->getRemainingFeatures();
Chris@0 321
Chris@0 322 for (size_t fi = 0; fi < features[m_outputFeatureNo].size(); ++fi) {
Chris@0 323 Vamp::Plugin::Feature feature =
Chris@0 324 features[m_outputFeatureNo][fi];
Chris@0 325 addFeature(blockFrame, feature);
Chris@0 326 }
Chris@0 327
Chris@0 328 if (frequencyDomain) {
Chris@0 329 for (size_t ch = 0; ch < channelCount; ++ch) {
Chris@3 330 delete fftModels[ch];
Chris@0 331 }
Chris@0 332 }
Chris@0 333
Chris@0 334 setCompletion(100);
Chris@0 335 }
Chris@0 336
Chris@0 337 void
Chris@0 338 FeatureExtractionPluginTransform::getFrames(int channel, int channelCount,
Chris@0 339 long startFrame, long size,
Chris@0 340 float *buffer)
Chris@0 341 {
Chris@0 342 long offset = 0;
Chris@0 343
Chris@0 344 if (startFrame < 0) {
Chris@0 345 for (int i = 0; i < size && startFrame + i < 0; ++i) {
Chris@0 346 buffer[i] = 0.0f;
Chris@0 347 }
Chris@0 348 offset = -startFrame;
Chris@0 349 size -= offset;
Chris@0 350 if (size <= 0) return;
Chris@0 351 startFrame = 0;
Chris@0 352 }
Chris@0 353
Chris@0 354 long got = getInput()->getValues
Chris@27 355 ((channelCount == 1 ? m_context.channel : channel),
Chris@0 356 startFrame, startFrame + size, buffer + offset);
Chris@0 357
Chris@0 358 while (got < size) {
Chris@0 359 buffer[offset + got] = 0.0;
Chris@0 360 ++got;
Chris@0 361 }
Chris@0 362
Chris@27 363 if (m_context.channel == -1 && channelCount == 1 &&
Chris@0 364 getInput()->getChannelCount() > 1) {
Chris@0 365 // use mean instead of sum, as plugin input
Chris@0 366 int cc = getInput()->getChannelCount();
Chris@0 367 for (long i = 0; i < size; ++i) {
Chris@0 368 buffer[i] /= cc;
Chris@0 369 }
Chris@0 370 }
Chris@0 371 }
Chris@0 372
Chris@0 373 void
Chris@0 374 FeatureExtractionPluginTransform::addFeature(size_t blockFrame,
Chris@0 375 const Vamp::Plugin::Feature &feature)
Chris@0 376 {
Chris@0 377 size_t inputRate = m_input->getSampleRate();
Chris@0 378
Chris@0 379 // std::cerr << "FeatureExtractionPluginTransform::addFeature("
Chris@0 380 // << blockFrame << ")" << std::endl;
Chris@0 381
Chris@0 382 int binCount = 1;
Chris@0 383 if (m_descriptor->hasFixedBinCount) {
Chris@0 384 binCount = m_descriptor->binCount;
Chris@0 385 }
Chris@0 386
Chris@0 387 size_t frame = blockFrame;
Chris@0 388
Chris@0 389 if (m_descriptor->sampleType ==
Chris@0 390 Vamp::Plugin::OutputDescriptor::VariableSampleRate) {
Chris@0 391
Chris@0 392 if (!feature.hasTimestamp) {
Chris@0 393 std::cerr
Chris@0 394 << "WARNING: FeatureExtractionPluginTransform::addFeature: "
Chris@0 395 << "Feature has variable sample rate but no timestamp!"
Chris@0 396 << std::endl;
Chris@0 397 return;
Chris@0 398 } else {
Chris@0 399 frame = Vamp::RealTime::realTime2Frame(feature.timestamp, inputRate);
Chris@0 400 }
Chris@0 401
Chris@0 402 } else if (m_descriptor->sampleType ==
Chris@0 403 Vamp::Plugin::OutputDescriptor::FixedSampleRate) {
Chris@0 404
Chris@0 405 if (feature.hasTimestamp) {
Chris@0 406 //!!! warning: sampleRate may be non-integral
Chris@0 407 frame = Vamp::RealTime::realTime2Frame(feature.timestamp,
Chris@0 408 m_descriptor->sampleRate);
Chris@0 409 } else {
Chris@0 410 frame = m_output->getEndFrame() + 1;
Chris@0 411 }
Chris@0 412 }
Chris@0 413
Chris@0 414 if (binCount == 0) {
Chris@0 415
Chris@0 416 SparseOneDimensionalModel *model = getOutput<SparseOneDimensionalModel>();
Chris@0 417 if (!model) return;
Chris@0 418 model->addPoint(SparseOneDimensionalModel::Point(frame, feature.label.c_str()));
Chris@0 419
Chris@0 420 } else if (binCount == 1) {
Chris@0 421
Chris@0 422 float value = 0.0;
Chris@0 423 if (feature.values.size() > 0) value = feature.values[0];
Chris@0 424
Chris@0 425 SparseTimeValueModel *model = getOutput<SparseTimeValueModel>();
Chris@0 426 if (!model) return;
Chris@0 427 model->addPoint(SparseTimeValueModel::Point(frame, value, feature.label.c_str()));
Chris@0 428
Chris@0 429 } else if (m_descriptor->sampleType ==
Chris@0 430 Vamp::Plugin::OutputDescriptor::VariableSampleRate) {
Chris@0 431
Chris@0 432 float pitch = 0.0;
Chris@0 433 if (feature.values.size() > 0) pitch = feature.values[0];
Chris@0 434
Chris@0 435 float duration = 1;
Chris@0 436 if (feature.values.size() > 1) duration = feature.values[1];
Chris@0 437
Chris@0 438 float velocity = 100;
Chris@0 439 if (feature.values.size() > 2) velocity = feature.values[2];
Chris@0 440
Chris@0 441 NoteModel *model = getOutput<NoteModel>();
Chris@0 442 if (!model) return;
Chris@0 443
Chris@0 444 model->addPoint(NoteModel::Point(frame, pitch, duration, feature.label.c_str()));
Chris@0 445
Chris@0 446 } else {
Chris@0 447
Chris@0 448 DenseThreeDimensionalModel::BinValueSet values = feature.values;
Chris@0 449
Chris@3 450 EditableDenseThreeDimensionalModel *model =
Chris@3 451 getOutput<EditableDenseThreeDimensionalModel>();
Chris@0 452 if (!model) return;
Chris@0 453
Chris@0 454 model->setBinValues(frame, values);
Chris@0 455 }
Chris@0 456 }
Chris@0 457
Chris@0 458 void
Chris@0 459 FeatureExtractionPluginTransform::setCompletion(int completion)
Chris@0 460 {
Chris@0 461 int binCount = 1;
Chris@0 462 if (m_descriptor->hasFixedBinCount) {
Chris@0 463 binCount = m_descriptor->binCount;
Chris@0 464 }
Chris@0 465
Chris@0 466 if (binCount == 0) {
Chris@0 467
Chris@0 468 SparseOneDimensionalModel *model = getOutput<SparseOneDimensionalModel>();
Chris@0 469 if (!model) return;
Chris@0 470 model->setCompletion(completion);
Chris@0 471
Chris@0 472 } else if (binCount == 1) {
Chris@0 473
Chris@0 474 SparseTimeValueModel *model = getOutput<SparseTimeValueModel>();
Chris@0 475 if (!model) return;
Chris@0 476 model->setCompletion(completion);
Chris@0 477
Chris@0 478 } else if (m_descriptor->sampleType ==
Chris@0 479 Vamp::Plugin::OutputDescriptor::VariableSampleRate) {
Chris@0 480
Chris@0 481 NoteModel *model = getOutput<NoteModel>();
Chris@0 482 if (!model) return;
Chris@0 483 model->setCompletion(completion);
Chris@0 484
Chris@0 485 } else {
Chris@0 486
Chris@3 487 EditableDenseThreeDimensionalModel *model =
Chris@3 488 getOutput<EditableDenseThreeDimensionalModel>();
Chris@0 489 if (!model) return;
Chris@0 490 model->setCompletion(completion);
Chris@0 491 }
Chris@0 492 }
Chris@0 493