// on a segment
void featureMapping(MixtureServer & ms, FeatureServer & fs,Seg * seg,Config &config) {
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName()); // Idx of the first frame of the current file in the feature server
fs.seekFeature(begin);
Feature f;
for (unsigned long idxFrame=0;idxFrame<seg->length();idxFrame++){ // for all the features of the segment
fs.readFeature(f,0);
featureMapping(ms,f,config);
fs.writeFeature(f);
}
}
RealVector <double> TopGauss::get(MixtureGD & UBM,FeatureServer &fs,Config & config){
RealVector <double> llkv;
for (unsigned long i=0;i<fs.getSourceCount();i++) {
String filename=fs.getNameOfASource(i);
if (verbose) cout << "(TopGauss::get on FeatureServer) File ["<<filename<<"] ... ";
this->read(filename,config); // this should be removed!! but when passing a feature server things get worse
llkv.addValue(get(UBM,fs,filename,config));
if (verbose) cout <<_nbgcnt/_nt <<" per frames LLK="<<llkv[i]<<endl;
}
return llkv;
}
// ----------------------------------------------------------------------------------------------------------
// Feature Warping giving a source(tab of histo, one by coeff) and a target distribution
// for a segment and cluster (segment is the minimum time unit to perform this)
void computeWarp(Histo *histoT,Histo &destH,FeatureServer & fs,unsigned long begin, unsigned long length,Config &config) {
unsigned long vectsize=fs.getVectSize(); // Get the vect size (number of coeff)
Feature f;
fs.seekFeature(begin);
for (unsigned long idxFrame=0;idxFrame<length;idxFrame++){ // for all the features of the segment
fs.readFeature(f,0);
// Get the feature;
for (unsigned int i = 0; i < vectsize; i++){ // For each coeff
f[i]=warping(f[i],histoT[i],destH); // Apply the warping function
}
fs.writeFeature(f);
}
}
// ----------------------------------------------------------------------------------------------------------
// Feature Mean subtraction and Cov reduction for a segment and cluster (segment is considerred to be the minimum time unit to perform this).
void computeZeroOne(const DoubleVector &featureMean,const DoubleVector &featureStd,FeatureServer & fs,unsigned long begin, unsigned long length,Config &config) {
unsigned long vectsize=fs.getVectSize(); // Get the vect size (number of coeff)
Feature f;
fs.seekFeature(begin);
for (unsigned long idxFrame=0;idxFrame<length;idxFrame++){ // for all the features of the segment
fs.readFeature(f,0);
// Get the feature;
for (unsigned int i = 0; i < vectsize; i++) { // For each coeff
f[i]=(f[i]-featureMean[i])/featureStd[i]; // Apply the 0 mean 1 cov normalisation
}
fs.writeFeature(f);
}
}
// Can use this function to get likelihood with a topgauss
double TopGauss::get(MixtureGD & UBM,FeatureServer &fs,String & featureFilename,Config & config){
StatServer ss(config);
String labelSelectedFrames =config.getParam("labelSelectedFrames");
unsigned long begin=fs.getFirstFeatureIndexOfASource(featureFilename);
fs.seekFeature(begin);
SegServer segmentsServer;
LabelServer labelServer;
initializeClusters(featureFilename,segmentsServer,labelServer,config);
// __android_log_print(ANDROID_LOG_DEBUG, "TopGauss::get", " Feature file %s \n", featureFilename.c_str());
verifyClusterFile(segmentsServer,fs,config);
unsigned long codeSelectedFrame=labelServer.getLabelIndexByString(labelSelectedFrames);
SegCluster& selectedSegments=segmentsServer.getCluster(codeSelectedFrame);
MixtureGDStat &acc=ss.createAndStoreMixtureStat(UBM);
Seg *seg; // current selected segment
selectedSegments.rewind();
unsigned long t=0; //cnt frames
acc.resetLLK();
unsigned long idxBegin=0;
while((seg=selectedSegments.getSeg())!=NULL){
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName());
fs.seekFeature(begin);
Feature f;
idxBegin=this->frameToIdx(t);
for (unsigned long idxFrame=0;idxFrame<seg->length();idxFrame++){
fs.readFeature(f);
//unsigned long idx=this->frameToIdx(t);
unsigned long nbg=_nbg[t];
ULongVector index;
double sumNonSelectedWeights=_snsw[t];
double sumNonSelectedLLK=_snsl[t];
for (unsigned long i=0;i<nbg;i++) {
index.addValue(_idx[idxBegin+i]);
}
char c[100];
sprintf(c,"%d",(int)index.size());
config.setParam("topDistribsCount",c); // this should be high enough
if (t==0) {acc.computeAndAccumulateLLK(f,1.0,DETERMINE_TOP_DISTRIBS);acc.resetLLK();} // to remove in ALIZE, this is to init the LKvector
ss.setTopDistribIndexVector(index, sumNonSelectedWeights, sumNonSelectedLLK);
acc.computeAndAccumulateLLK(f,1.0,USE_TOP_DISTRIBS);
idxBegin+=nbg;
t++;
}
}
//ss.deleteMixtureStat(acc);
if (t!=_nt || idxBegin !=_nbgcnt) cout << "W: t("<<t<<") != _nt(" <<_nt<<")"<<"W: idxBegin("<<idxBegin<<") != _nbgcnt(" <<_nbgcnt<<")"<<endl;
return acc.getMeanLLK();
}
// on a part of a file
void outputFeatureFile(Config &config, FeatureServer &fs, unsigned long begin,unsigned long length, FeatureFileWriter &w) {
Feature f;
fs.seekFeature(begin);
for (unsigned long idxFrame=0;idxFrame<length;idxFrame++){
outputFeatureFile(config, fs, f,w);
}
}
void FactorAnalysisStat::computeAndAccumulateGeneralFAStats(SegCluster &selectedSegments,FeatureServer &fs,Config & config){
if (verbose) cout <<"(FactorAnalysisStat) Compute General FA Stats (Complete)" << endl;
double *N_h, *N, *S_X_h, *S_X,*ff;
_matN_h.setAllValues(0.0);
_matN.setAllValues(0.0);
_matS_X_h.setAllValues(0.0);
_matS_X.setAllValues(0.0);
N_h=_matN_h.getArray(); N=_matN.getArray(); S_X_h=_matS_X_h.getArray();S_X=_matS_X.getArray();
MixtureGD & UBM=_ms.getMixtureGD((unsigned long) 1);
MixtureGDStat &acc=_ss.createAndStoreMixtureStat(UBM);
// Compute Occupations and Statistics
acc.resetOcc();
Seg *seg;
selectedSegments.rewind();
String currentSource="";unsigned long loc=0;unsigned long sent=0;
while((seg=selectedSegments.getSeg())!=NULL){
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName()); // Idx of the first frame of the current file in the feature server
if (currentSource!=seg->sourceName()) {
currentSource=seg->sourceName();
loc=_ndxTable.locNb(currentSource);
sent=_ndxTable.sessionNb(currentSource);
if (verbose)cout << "Processing speaker["<<currentSource<<"]"<< endl;
}
fs.seekFeature(begin);
Feature f;
if (!_topGauss) {
for (unsigned long idxFrame=0;idxFrame<seg->length();idxFrame++){
fs.readFeature(f);
acc.computeAndAccumulateOcc(f);
RealVector <double> aPost=acc.getOccVect();
ff=f.getDataVector();
for(unsigned long k=0;k<_mixsize;k++) {
N_h[sent*_mixsize+k]+=aPost[k];
N[loc*_mixsize+k] +=aPost[k];
for (unsigned long i=0;i<_vsize;i++) {
S_X_h[sent*_supervsize+(k*_vsize+i)]+=aPost[k]*ff[i];
S_X[loc*_supervsize+(k*_vsize+i)] +=aPost[k]*ff[i];
}
}
}
}
else throw Exception("ComputeGeneralStats TopGauss not done at this level",__FILE__,__LINE__);
}
};
/// Normalize features with a smooth mixture transformation o't=ot-sum(P(c|ot)Uc.x)
void FactorAnalysisStat::normalizeFeatures(SegCluster &selectedSegments,FeatureServer &fs,Config & config){
if (verbose) cout << "(FactorAnalysisStat) Normalize Features" << endl;
MixtureGD & clientMixture=_ms.getMixtureGD(1); // copy the UBM mixture
unsigned long nt=0;
RealVector <double> m_xh_1; m_xh_1.setSize(_supervsize);
double *_m_xh_1=m_xh_1.getArray();
Seg *seg; // current selectd segment
selectedSegments.rewind();
String currentSource="";
while((seg=selectedSegments.getSeg())!=NULL){
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName());
if (currentSource!=seg->sourceName()) {
currentSource=seg->sourceName();
this->getUX(m_xh_1,currentSource);
this->getSpeakerModel(clientMixture,currentSource);
if (verbose)cout << "Processing speaker["<<currentSource<<"]"<< endl;
}
fs.seekFeature(begin);
Feature f;
if (!_topGauss) {
for (unsigned long idxFrame=0;idxFrame<seg->length();idxFrame++){
fs.readFeature(f,0);
double *ff=f.getDataVector();
double sum=0.0;
RealVector <double> P;
P.setSize(_mixsize);
double *Prob=P.getArray();
for(unsigned long k=0;k<_mixsize;k++) {
Prob[k]=clientMixture.weight(k)*clientMixture.getDistrib(k).computeLK(f);
sum+=Prob[k];
}
for(unsigned long k=0;k<_mixsize;k++)
Prob[k]/=sum;
for(unsigned long k=0;k<_mixsize;k++) {
for (unsigned long i=0;i<_vsize;i++)
ff[i]-= Prob[k]*_m_xh_1[k*_vsize+i];
}
fs.writeFeature(f);
nt++;
}
}
else {
throw Exception("no topgauss yet",__FILE__,__LINE__);
}
}
};
void cms(String & featureFileName,FeatureServer &fs,Config &config) {
unsigned long begin=fs.getFirstFeatureIndexOfASource(featureFileName);
fs.seekFeature(begin);
SegServer segmentsServer;
LabelServer labelServer;
initializeClusters(featureFileName,segmentsServer,labelServer,config);
verifyClusterFile(segmentsServer,fs,config);
unsigned long codeSelectedFrame=labelServer.getLabelIndexByString(config.getParam("labelSelectedFrames"));
SegCluster& selectedSegments=segmentsServer.getCluster(codeSelectedFrame);
selectedSegments.rewind();
RealVector <double> mean,cov;
FrameAccGD frameAccu;
frameAccu.reset();
accumulateStatFrame(frameAccu,fs, selectedSegments, config);
mean = frameAccu.getMeanVect(); // Get the mean vector
cov = frameAccu.getStdVect(); // Get the std vector
computeZeroOne(mean,cov,fs, selectedSegments, config);
}
/// Compute Log Likelihood of the Factor Analysis model
double FactorAnalysisStat::getLLK(SegCluster &selectedSegments,MixtureGD &model,FeatureServer&fs,Config & config){
if (verbose) cout << "(FactorAnalysisStat) Compute Likelihood" << endl;
double llk=0.0;
MixtureGDStat &acc=_ss.createAndStoreMixtureStat(model);
Seg *seg;
selectedSegments.rewind();
while((seg=selectedSegments.getSeg())!=NULL){
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName());
fs.seekFeature(begin);
Feature f;
for (unsigned long idxFrame=0;idxFrame<seg->length();idxFrame++){
fs.readFeature(f);
acc.computeAndAccumulateLLK(f,1.0,TOP_DISTRIBS_NO_ACTION);
}
}
llk= acc.getMeanLLK();
_ss.deleteMixtureStat(acc);
return llk;
};
double computeEnergyThreshold(FeatureServer & fs,double pSelect,unsigned long nbBins=100)
{
Histo histo(nbBins); // Create an histo accumulator with 100 bins
Feature f; // reset the reader at the begin of the input stream
fs.reset(); // feature server reset
for (unsigned long ind=0;fs.readFeature(f); ind++) // feature loop
histo.accumulateValue(f[0]); // Accumulate the energy in the histo Accumulator
histo.computeHisto(); // Compute the histo
long i=nbBins-1;
real_t count=0;
while((i>=0) && (count<=pSelect)){ // Find the bin corresponding to the percentage of data wanted
count+=histo.count(i)*(histo.higherBound(i)-histo.lowerBound(i));
i--;
}
double threshold;
if (i>=0) threshold=histo.higherBound(i); // Set the threshold to the higherBound of the next bin
else threshold=histo.lowerBound(0); // if 100% of data should be selected
if (verbose) cout << "Percentage wanted["<<(int) (pSelect*100.0) <<"]Energy threshold["<<threshold<<"]"<<endl;
return threshold;
}
// Build the segments with the energized frames
unsigned long selectFrames(FeatureServer &fs,SegServer & segServer,double threshold,SegCluster &selectedSeg,SegCluster &outputSeg,String labelOutput,String fileName)
{
unsigned long countFrames=0;
fs.reset(); // feature server reset
unsigned long ind=0;
unsigned long begin=0;
bool in=false;
Seg *seg; // current selectd segment
selectedSeg.rewind(); // reset the reader at the begin of the input stream
while((seg=selectedSeg.getSeg())!=NULL){ // For each input segments
for (unsigned long idx=seg->begin();idx<seg->begin()+seg->length();idx++){ // for each frame
Feature f;
fs.seekFeature(idx);
fs.readFeature(f);
if (f[0]>threshold){ // the frame is selected
countFrames++;
if (in==false){ // Begin of a new segment
in=true;
begin=ind;
}
}
else if (in){ // End of a segment
in=false;
Seg & segFake=segServer.createSeg(begin,ind-begin,0, // Create a segment - Take care : length=end-begin+1 but ind =end+1 !!
labelOutput,fileName);
outputSeg.add(segFake); // Add a segment
}
ind++; // Increment the frame index
} // end of one input segment
if (in){ // deal with the last energized segmeent inside the current input segment
in=false;
Seg & segFake=segServer.createSeg(begin,ind-begin+1,0, // Create a segment
labelOutput,fileName);
outputSeg.add(segFake); // Add a segment - Take care : length=end-begin+1 and ind=end in this case !!
}
} // end feature loop
return countFrames;
}
// Main init function
double TopGauss::compute(MixtureGD & UBM,FeatureServer &fs,String & featureFilename,Config & config){
StatServer ss(config);
MixtureGDStat &acc=ss.createAndStoreMixtureStat(UBM);
unsigned long _mixsize=UBM.getDistribCount();
String labelSelectedFrames =config.getParam("labelSelectedFrames");
unsigned long begin=fs.getFirstFeatureIndexOfASource(featureFilename);
fs.seekFeature(begin);
SegServer segmentsServer;
LabelServer labelServer;
initializeClusters(featureFilename,segmentsServer,labelServer,config);
// __android_log_print(ANDROID_LOG_DEBUG, "TopGauss::compute", " Feature file %s \n", featureFilename.c_str());
verifyClusterFile(segmentsServer,fs,config);
unsigned long codeSelectedFrame=labelServer.getLabelIndexByString(labelSelectedFrames);
SegCluster& selectedSegments=segmentsServer.getCluster(codeSelectedFrame);
acc.resetLLK();
double topD=config.getParam("topGauss").toDouble();
if (verbose) {if(topD<1.0) cout << "LLK %="<< topD << "% ";else cout << "Top-"<<topD<<" ";}
// Class values
_nt=totalFrame(selectedSegments);
_nbg.setSize(_nt); _idx.setSize(0);_snsw.setSize(0); _snsl.setSize(0);
_nbg.setAllValues(0); _idx.setAllValues(0);_snsw.setAllValues(0.0);_snsl.setAllValues(0.0);
_nbgcnt=0;
Seg *seg; // current selected segment
selectedSegments.rewind();
unsigned long t=0; //cnt frames
while((seg=selectedSegments.getSeg())!=NULL){
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName());
fs.seekFeature(begin);
Feature f;
for (unsigned long idxFrame=0;idxFrame<seg->length();idxFrame++){
fs.readFeature(f);
double llk=acc.computeAndAccumulateLLK(f,1.0,DETERMINE_TOP_DISTRIBS);
const LKVector &topV=ss.getTopDistribIndexVector();
double lk_tot=exp(llk);
double val=0.0;
if (topD<1.0) {
for(unsigned long j=0;j<_mixsize;j++){
if (val > topD*lk_tot) break;
val+=(topV[j].lk);
_nbg[t]++;
}
} else _nbg[t]=(unsigned long)topD;
_nbgcnt+=_nbg[t];
double snsw=1.0;
double snsl=lk_tot;
for(unsigned long j=0;j<_nbg[t];j++) {
_idx.addValue(topV[j].idx);
snsw -=UBM.weight(topV[j].idx);
snsl -=topV[j].lk;
}
_snsw.addValue(snsw);
if (snsl < EPS_LK)
_snsl.addValue(EPS_LK);
else _snsl.addValue(snsl);
t++;
}
}
if (t!=_nt) cout << "W: t("<<t<<") != _nt(" <<_nt<<")"<<endl;
return acc.getMeanLLK();
}
// on a segment
void outputFeatureFile(Config &config, FeatureServer &fs, Seg * seg, FeatureFileWriter &w) {
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName()); // Idx of the first frame of the current file in the feature server
if (verbose) cout <<"(GeneralTools) Writing ["<<seg->sourceName()<<"]"<<" begin:"<<begin<<" length:"<<seg->length()<<endl;
outputFeatureFile(config,fs,begin,seg->length(),w);
}
// Feature writing in an output stream w - could be used for multiple segmen,ts from multiple files to one file
void outputFeatureFile(Config &config, FeatureServer &fs, Feature & f, FeatureFileWriter &w) {
fs.readFeature(f);
w.writeFeature(f);
}
void computeZeroOne(const DoubleVector &featureMean,const DoubleVector &featureStd, FeatureServer & fs, Seg* seg, Config & config){
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName()); // Idx of the first frame of the current file in the feature server
computeZeroOne(featureMean,featureStd,fs,begin,seg->length(),config); // Normalize the feature to fit 0 mean, 1 cov
}
// on a segment
void computeZeroOne(FrameAccGD & frameAccu, FeatureServer & fs, Seg* seg, Config & config){
const DoubleVector & featureMean = frameAccu.getMeanVect(); // Get the mean vector
const DoubleVector & featureStd = frameAccu.getStdVect(); // Get the std vector (sqrt(cov))
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName()); // Idx of the first frame of the current file in the feature server
computeZeroOne(featureMean,featureStd,fs,begin,seg->length(),config); // Normalize the feature to fit 0 mean, 1 cov
}
// on a segment
void computeWarp(Histo *histoT,Histo &destH, FeatureServer & fs, Seg* seg,Config & config){
unsigned long begin=seg->begin()+fs.getFirstFeatureIndexOfASource(seg->sourceName()); // Idx of the first frame of the current file in the feature server
computeWarp(histoT,destH,fs,begin,seg->length(),config);
}
// On a complete feature stream
void globalMeanCov (FeatureServer &fs,FrameAcc & globalFrameAcc,Config &config) {
globalFrameAcc.reset();
accumulateStatFrame(globalFrameAcc,fs,0,fs.getFeatureCount(),config);
}