Mentions légales du service

Skip to content
Snippets Groups Projects
Commit c800223b authored by MONSEIGNE Thibaut's avatar MONSEIGNE Thibaut
Browse files

Fix : Find Regularized CSP Bug

parent 01514a55
Branches
Tags
No related merge requests found
Hide whitespace changes
Inline Side-by-side
plugins/processing/signal-processing/src/box-algorithms/filters/ovpCBoxAlgorithmRegularizedCSPTrainer.cpp
+ 24
− 40
View file @ c800223b
......@@ -11,23 +11,19 @@
using namespace OpenViBE;
using namespace /*OpenViBE::*/Kernel;
using namespace /*OpenViBE::*/Plugins;
using namespace SignalProcessing;
using namespace /*OpenViBE::Plugins::*/SignalProcessing;
using namespace Eigen;
using namespace std;
// typedef Eigen::Matrix< double, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor > MatrixXdRowMajor;
CBoxAlgorithmRegularizedCSPTrainer::CBoxAlgorithmRegularizedCSPTrainer() : m_tikhonov(0.0) {}
bool CBoxAlgorithmRegularizedCSPTrainer::initialize()
{
m_stimDecoder.initialize(*this, 0);
m_encoder.initialize(*this, 0);
const IBox& staticBoxContext = this->getStaticBoxContext();
m_nClasses = staticBoxContext.getInputCount() - 1;
m_nClasses = this->getStaticBoxContext().getInputCount() - 1;
m_covProxies.resize(m_nClasses);
......@@ -36,7 +32,6 @@ bool CBoxAlgorithmRegularizedCSPTrainer::initialize()
m_filtersPerClass = size_t(uint64_t(FSettingValueAutoCast(*this->getBoxAlgorithmContext(), 2)));
m_saveAsBoxConf = FSettingValueAutoCast(*this->getBoxAlgorithmContext(), 3);
// @FIXME CERT
OV_ERROR_UNLESS_KRF(m_filtersPerClass > 0, // && m_filtersPerClass%2 == 0,
"Invalid filter dimension number [" << m_filtersPerClass << "] (expected value > 0)", // even ?
ErrorType::BadSetting);
......@@ -49,14 +44,10 @@ bool CBoxAlgorithmRegularizedCSPTrainer::initialize()
m_signalDecoders[i].initialize(*this, i + 1);
const CIdentifier covAlgId = this->getAlgorithmManager().createAlgorithm(OVP_ClassId_Algorithm_OnlineCovariance);
OV_ERROR_UNLESS_KRF(covAlgId != OV_UndefinedIdentifier, "Failed to create online covariance algorithm",
ErrorType::BadResourceCreation);
OV_ERROR_UNLESS_KRF(covAlgId != OV_UndefinedIdentifier, "Failed to create online covariance algorithm", ErrorType::BadResourceCreation);
m_covProxies[i].cov = &this->getAlgorithmManager().getAlgorithm(covAlgId);
OV_ERROR_UNLESS_KRF(m_covProxies[i].cov->initialize(), "Failed to initialize online covariance algorithm",
ErrorType::Internal);
OV_ERROR_UNLESS_KRF(m_covProxies[i].cov->initialize(), "Failed to initialize online covariance algorithm", ErrorType::Internal);
// Set the params of the cov algorithm
TParameterHandler<uint64_t> updateMethod(m_covProxies[i].cov->getInputParameter(OVP_Algorithm_OnlineCovariance_InputParameterId_UpdateMethod));
......@@ -70,8 +61,7 @@ bool CBoxAlgorithmRegularizedCSPTrainer::initialize()
m_tikhonov = FSettingValueAutoCast(*this->getBoxAlgorithmContext(), 7);
OV_ERROR_UNLESS_KRF(m_configFilename != CString(""), "Output filename is required in box configuration",
ErrorType::BadSetting);
OV_ERROR_UNLESS_KRF(m_configFilename != CString(""), "Output filename is required in box configuration", ErrorType::BadSetting);
return true;
}
......@@ -106,9 +96,9 @@ bool CBoxAlgorithmRegularizedCSPTrainer::processInput(const size_t /*index*/)
bool CBoxAlgorithmRegularizedCSPTrainer::updateCov(const size_t index)
{
IBoxIO& dynamicBoxContext = this->getDynamicBoxContext();
IBoxIO& boxCtx = this->getDynamicBoxContext();
SIncrementalCovarianceProxy& curCovProxy(m_covProxies[index]);
for (size_t i = 0; i < dynamicBoxContext.getInputChunkCount(index + 1); ++i)
for (size_t i = 0; i < boxCtx.getInputChunkCount(index + 1); ++i)
{
auto* decoder = &m_signalDecoders[index];
const IMatrix* matrix = decoder->getOutputMatrix();
......@@ -133,14 +123,14 @@ bool CBoxAlgorithmRegularizedCSPTrainer::updateCov(const size_t index)
}
if (decoder->isBufferReceived())
{
TParameterHandler<IMatrix*> ip_features(curCovProxy.cov->getInputParameter(OVP_Algorithm_OnlineCovariance_InputParameterId_InputVectors));
TParameterHandler<IMatrix*> features(curCovProxy.cov->getInputParameter(OVP_Algorithm_OnlineCovariance_InputParameterId_InputVectors));
// transpose data
const size_t nChannels = matrix->getDimensionSize(0);
const size_t nSamples = matrix->getDimensionSize(1);
const Map<MatrixXdRowMajor> inputMapper(const_cast<double*>(matrix->getBuffer()), nChannels, nSamples);
Map<MatrixXdRowMajor> outputMapper(ip_features->getBuffer(), nSamples, nChannels);
Map<MatrixXdRowMajor> outputMapper(features->getBuffer(), nSamples, nChannels);
outputMapper = inputMapper.transpose();
curCovProxy.cov->activateInputTrigger(OVP_Algorithm_OnlineCovariance_Process_Update, true);
......@@ -149,12 +139,8 @@ bool CBoxAlgorithmRegularizedCSPTrainer::updateCov(const size_t index)
curCovProxy.nBuffers++;
curCovProxy.nSamples += nSamples;
}
if (decoder->isEndReceived())
{
// nop
}
// if (decoder->isEndReceived()) { } // nop
}
return true;
}
......@@ -207,14 +193,11 @@ bool CBoxAlgorithmRegularizedCSPTrainer::computeCSP(const vector<MatrixXd>& cov,
sortedEigenVectors.resize(m_nClasses);
sortedEigenValues.resize(m_nClasses);
// To get the CSP filters, we compute two sets of eigenvectors,
// eig(inv(sigma2+tikhonov)*sigma1) and eig(inv(sigma1+tikhonov)*sigma2
// and pick the ones corresponding to the largest eigenvalues as
// spatial filters [following Lotte & Guan 2011]. Assumes the shrink
// of the sigmas (if its used) has been performed inside the cov
// computation algorithm.
// To get the CSP filters, we compute two sets of eigenvectors, eig(inv(sigma2+tikhonov)*sigma1) and eig(inv(sigma1+tikhonov)*sigma2
// and pick the ones corresponding to the largest eigenvalues as spatial filters [following Lotte & Guan 2011]. Assumes the shrink
// of the sigmas (if its used) has been performed inside the cov computation algorithm.
EigenSolver<MatrixXd> eigenSolverGeneral;
EigenSolver<MatrixXd> solver;
for (size_t c = 0; c < m_nClasses; ++c)
{
......@@ -227,11 +210,11 @@ bool CBoxAlgorithmRegularizedCSPTrainer::computeCSP(const vector<MatrixXd>& cov,
covProd[c] = covInv[c] * outclassCov;
try { eigenSolverGeneral.compute(covProd[c]); }
try { solver.compute(covProd[c]); }
catch (...) { OV_ERROR_KRF("EigenSolver failed for condition [" << c + 1 << "]", ErrorType::BadProcessing); }
eigenValues[c] = eigenSolverGeneral.eigenvalues().real();
eigenVectors[c] = eigenSolverGeneral.eigenvectors().real();
eigenValues[c] = solver.eigenvalues().real();
eigenVectors[c] = solver.eigenvectors().real();
// Sort the vectors -_-
vector<pair<double, int>> indexes;
......@@ -241,9 +224,11 @@ bool CBoxAlgorithmRegularizedCSPTrainer::computeCSP(const vector<MatrixXd>& cov,
sortedEigenValues[c].resizeLike(eigenValues[c]);
sortedEigenVectors[c].resizeLike(eigenVectors[c]);
for (int i = 0; i < eigenValues[c].size(); ++i)
{
sortedEigenValues[c][i] = eigenValues[c][indexes[i].second];
//@todo @FIXME This fonction work sometimes randomly
sortedEigenVectors[c].col(i) = eigenVectors[c].col(indexes[i].second);
}
}
......@@ -336,9 +321,9 @@ bool CBoxAlgorithmRegularizedCSPTrainer::process()
}
// Compute the actual CSP using the obtained covariance matrices
vector<MatrixXd> sortedEigenVectors;
vector<VectorXd> sortedEigenValues;
OV_ERROR_UNLESS_KRF(computeCSP(cov, sortedEigenVectors, sortedEigenValues), "Failure when computing CSP", ErrorType::BadProcessing);
vector<MatrixXd> sortedVectors;
vector<VectorXd> sortedValues;
OV_ERROR_UNLESS_KRF(computeCSP(cov, sortedVectors, sortedValues), "Failure when computing CSP", ErrorType::BadProcessing);
// Create a CMatrix mapper that can spool the filters to a file
CMatrix selectedVectors;
......@@ -350,11 +335,10 @@ bool CBoxAlgorithmRegularizedCSPTrainer::process()
for (size_t c = 0; c < m_nClasses; ++c)
{
selectedVectorsMapper.block(c * m_filtersPerClass, 0, m_filtersPerClass, nChannels) = sortedEigenVectors[c]
.block(0, 0, nChannels, m_filtersPerClass).transpose();
selectedVectorsMapper.block(c * m_filtersPerClass, 0, m_filtersPerClass, nChannels) = sortedVectors[c].block(0, 0, nChannels, m_filtersPerClass).transpose();
this->getLogManager() << LogLevel_Info << "The " << m_filtersPerClass << " filter(s) for cond " << c + 1 << " cover "
<< 100.0 * sortedEigenValues[c].head(m_filtersPerClass).sum() / sortedEigenValues[c].sum() << "% of corresp. eigenvalues\n";
<< 100.0 * sortedValues[c].head(m_filtersPerClass).sum() / sortedValues[c].sum() << "% of corresp. eigenvalues\n";
}
if (m_saveAsBoxConf)
......
plugins/processing/signal-processing/src/box-algorithms/filters/ovpCBoxAlgorithmRegularizedCSPTrainer.h
+ 1
− 2
View file @ c800223b
......@@ -21,7 +21,6 @@ namespace OpenViBE
class CBoxAlgorithmRegularizedCSPTrainer final : public Toolkit::TBoxAlgorithm<IBoxAlgorithm>
{
public:
CBoxAlgorithmRegularizedCSPTrainer();
void release() override { delete this; }
bool initialize() override;
bool uninitialize() override;
......@@ -45,7 +44,7 @@ namespace OpenViBE
bool m_saveAsBoxConf = false;
bool m_hasBeenInitialized = false;
double m_tikhonov = 0;
double m_tikhonov = 0.0;
struct SIncrementalCovarianceProxy
{
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment