// accumulate statistics
void DTAccumulator::accumulate(Alignment *alignment, MatrixBase<float> &mFeatures, bool bNumerator) {
Accumulator *accumulator = NULL;
double dOccupationTotal = 0.0;
for(unsigned int t=0 ; t < alignment->getFrames() ; ++t) {
FrameAlignment *frameAlignment = alignment->getFrameAlignment(t);
VectorStatic<float> vFeatureVector = mFeatures.getRow(t);
for(FrameAlignment::iterator it = frameAlignment->begin() ; it != frameAlignment->end() ; ++it) {
double dOccupationNum = (*it)->dOccupation;
HMMState *hmmState = m_hmmManager->getHMMState((*it)->iHMMState);
// get Gaussian occupation from the mixture occupation
// (1) compute the mixture likelihood (all Gaussian components)
double dLikelihoodTotal = -DBL_MAX;
int iGaussianComponents = hmmState->getMixture().getNumberComponents();
double *dLikelihoodGaussian = new double[iGaussianComponents];
for(int iGaussian = 0 ; iGaussian < iGaussianComponents ; ++iGaussian) {
dLikelihoodGaussian[iGaussian] = hmmState->computeEmissionProbabilityGaussian(iGaussian,
mFeatures.getRow(t).getData(),-1);
dLikelihoodGaussian[iGaussian] += log(hmmState->getMixture()(iGaussian)->weight());
dLikelihoodTotal = Numeric::logAddition(dLikelihoodTotal,dLikelihoodGaussian[iGaussian]);
}
// (2) accumulate statistics for each mixture component
for(int iGaussian = 0 ; iGaussian < iGaussianComponents ; ++iGaussian) {
double dProbGaussian = exp(dLikelihoodGaussian[iGaussian]-dLikelihoodTotal);
assert(dProbGaussian >= 0.0);
double dOccupationGaussian = dOccupationNum*dProbGaussian;
unsigned int iKey = Accumulator::getPhysicalAccumulatorKey(hmmState->getId(),iGaussian);
if (bNumerator) {
accumulator = m_mAccumulatorNum[iKey];
} else {
accumulator = m_mAccumulatorDen[iKey];
}
accumulator->accumulateObservation(vFeatureVector,dOccupationGaussian);
dOccupationTotal += dOccupationGaussian;
}
delete [] dLikelihoodGaussian;
}
}
printf("accumulated: %12.6f\n",dOccupationTotal);
}
// feed adaptation data from an alignment
void MLLRManager::feedAdaptationData(MatrixBase<float> &mFeatures, Alignment *alignment, double *dLikelihood) {
// sanity check
unsigned int iFeatures = mFeatures.getRows();
assert(iFeatures == alignment->getFrames());
m_iAdaptationFrames += iFeatures;
*dLikelihood = 0.0;
for(unsigned int t=0 ; t<iFeatures ; ++t) {
FrameAlignment *frameAlignment = alignment->getFrameAlignment(t);
VectorStatic<float> vFeatureVector = mFeatures.getRow(t);
for(FrameAlignment::iterator it = frameAlignment->begin() ; it != frameAlignment->end() ; ++it) {
HMMStateDecoding *hmmStateDecoding = m_hmmManager->getHMMStateDecoding((*it)->iHMMState);
// compute the contribution of each Gaussian
// (case 1) all the frame-level adaptation data goes to the best scoring Gaussian component (faster)
if (m_bBestComponentOnly) {
float fLikelihood = -FLT_MAX;
GaussianDecoding *gaussian = hmmStateDecoding->getBestScoringGaussian(vFeatureVector.getData(),&fLikelihood);
*dLikelihood += std::max<float>(fLikelihood,LOG_LIKELIHOOD_FLOOR);
GaussianStats *gaussianStats = NULL;
if (m_gaussianStats[gaussian->iId] == NULL) {
gaussianStats = new GaussianStats;
gaussianStats->gaussian = gaussian;
gaussianStats->dOccupation = 0.0;
gaussianStats->vObservation = new Vector<double>(m_iDim);
gaussianStats->vObservation->zero();
m_gaussianStats[gaussian->iId] = gaussianStats;
} else {
gaussianStats = m_gaussianStats[gaussian->iId];
}
gaussianStats->dOccupation += 1.0;
gaussianStats->vObservation->add(1.0,vFeatureVector);
m_regressionTree->accumulateStatistics(vFeatureVector.getData(),1.0,gaussianStats);
}
// (case 2) adaptation data is shared across all components (slightly more accurate)
else {
double *dProbGaussian = new double[hmmStateDecoding->getGaussianComponents()];
double dProbTotal = 0.0;
for(int g=0 ; g < hmmStateDecoding->getGaussianComponents() ; ++g) {
dProbGaussian[g] = exp(hmmStateDecoding->computeGaussianProbability(g,vFeatureVector.getData()));
dProbTotal += dProbGaussian[g];
assert((finite(dProbGaussian[g])) && (finite(dProbTotal)));
}
*dLikelihood += max(log(dProbTotal),LOG_LIKELIHOOD_FLOOR);
for(int g=0 ; g < hmmStateDecoding->getGaussianComponents() ; ++g) {
dProbGaussian[g] /= dProbTotal;
double dOccupation = dProbGaussian[g]*1.0;
GaussianDecoding *gaussian = hmmStateDecoding->getGaussian(g);
GaussianStats *gaussianStats = NULL;
if (m_gaussianStats[gaussian->iId] == NULL) {
gaussianStats = new GaussianStats;
gaussianStats->gaussian = gaussian;
gaussianStats->dOccupation = 0.0;
gaussianStats->vObservation = new Vector<double>(m_iDim);
gaussianStats->vObservation->zero();
m_gaussianStats[gaussian->iId] = gaussianStats;
} else {
gaussianStats = m_gaussianStats[gaussian->iId];
}
gaussianStats->dOccupation += dOccupation;
gaussianStats->vObservation->add(dOccupation,vFeatureVector);
m_regressionTree->accumulateStatistics(vFeatureVector.getData(),dOccupation,gaussianStats);
}
delete [] dProbGaussian;
}
}
}
}
// feed adaptation data from an alignment into the adaptation process
// note: it is possible to accumulate statistics at the Gaussian component or at the transform level
// - Gaussian component level: O(n^2) per Gaussian component
// - Transform level: O(n^3) per transform + O(n) per Gaussian component
void FMLLREstimator::feedAdaptationData(MatrixBase<float> &mFeatures, Alignment *alignment, double *dLikelihood) {
// sanity check
assert((unsigned int)mFeatures.getRows() == alignment->getFrames());
m_fOccupancyTotal += mFeatures.getRows();
*dLikelihood = 0.0;
for(unsigned int t=0 ; t<mFeatures.getRows() ; ++t) {
VectorStatic<float> vFeatureVector = mFeatures.getRow(t);
// extended observation vector
Vector<double> vObsEx(vFeatureVector);
vObsEx.appendFront(1.0);
// for each HMM-state the observation is assigned to
FrameAlignment *frameAlignment = alignment->getFrameAlignment(t);
for(FrameAlignment::iterator it = frameAlignment->begin() ; it != frameAlignment->end() ; ++it) {
HMMStateDecoding *hmmStateDecoding = m_hmmManager->getHMMStateDecoding((*it)->iHMMState);
// compute the contribution of each Gaussian
// (case 1) all the frame-level adaptation data goes to the best scoring Gaussian component (faster)
if (m_bBestComponentOnly) {
float fLikelihood = -FLT_MAX;
GaussianDecoding *gaussian = hmmStateDecoding->getBestScoringGaussian(vFeatureVector.getData(),&fLikelihood);
*dLikelihood += std::max<float>(fLikelihood,LOG_LIKELIHOOD_FLOOR);
Vector<float> vCovariance(VectorStatic<float>(gaussian->fCovariance,m_iDim));
#ifdef OPTIMIZED_COMPUTATION
vCovariance.mul(2.0);
vCovariance.invertElements();
#endif
// accumulate data for each G(i)
m_matrixAux->zero();
m_matrixAux->addVecMul(1.0,vObsEx,vObsEx);
for(int i=0 ; i < m_iDim; ++i) {
double dCovInv = 1.0/vCovariance(i);
m_matrixG[i]->add(dCovInv*1.0,*m_matrixAux);
}
// accumulate data for each k(i)
for(int i=0 ; i < m_iDim ; ++i) {
double dConstant = (gaussian->fMean[i]/vCovariance(i))*1.0;
m_matrixK->getRow(i).add(dConstant,vObsEx);
}
}
// (case 2) adaptation data is shared across all components (slightly more accurate)
else {
double *dProbGaussian = new double[hmmStateDecoding->getGaussianComponents()];
double dProbTotal = 0.0;
for(int g=0 ; g < hmmStateDecoding->getGaussianComponents() ; ++g) {
dProbGaussian[g] = exp(hmmStateDecoding->computeGaussianProbability(g,vFeatureVector.getData()));
dProbTotal += dProbGaussian[g];
assert((finite(dProbGaussian[g])) && (finite(dProbTotal)));
}
*dLikelihood += max(log(dProbTotal),LOG_LIKELIHOOD_FLOOR);
for(unsigned int iGaussian = 0 ; iGaussian < hmmStateDecoding->getMixtureSize() ; ++iGaussian) {
GaussianDecoding *gaussian = hmmStateDecoding->getGaussian(iGaussian);
double dGaussianOccupation = dProbGaussian[iGaussian]/dProbTotal;
Vector<float> vCovariance(VectorStatic<float>(gaussian->fCovariance,m_iDim));
#ifdef OPTIMIZED_COMPUTATION
vCovariance.mul(2.0);
vCovariance.invertElements();
#endif
// accumulate data for each G(i)
m_matrixAux->zero();
m_matrixAux->addVecMul(1.0,vObsEx,vObsEx);
for(int i=0 ; i < m_iDim; ++i) {
double dCovInv = 1.0/vCovariance(i);
m_matrixG[i]->add(dCovInv*dGaussianOccupation,*m_matrixAux);
}
// accumulate data for each k(i)
for(int i=0 ; i < m_iDim ; ++i) {
double dConstant = (gaussian->fMean[i]/vCovariance(i))*dGaussianOccupation;
m_matrixK->getRow(i).add(dConstant,vObsEx);
}
}
delete [] dProbGaussian;
}
}
}
}
