WHcoord fileManager::meanCoordFromMask() const
{
if (m_maskMatrix.empty()) {
std::cerr<< "ERROR @ fileManager::storeFullTract(): Mask hast not been loaded, returning 0 coordinate"<<std::endl;
return WHcoord();
}
size_t sumX( 0 ), sumY( 0 ), sumZ( 0 ), sumElements( 0 );
for( int i=0 ; i<m_maskMatrix.size() ; ++i )
{
for( int j=0 ; j< m_maskMatrix[i].size() ; ++j )
{
for( int k=0 ; k<m_maskMatrix[i][j].size() ; ++k )
{
if (m_maskMatrix[i][j][k])
{
sumX += i;
sumY += j;
sumZ += k;
++sumElements;
}
}
}
}
size_t meanX( sumX/sumElements );
size_t meanY( sumY/sumElements );
size_t meanZ( sumZ/sumElements );
WHcoord meanCoord(meanX,meanY,meanZ);
return meanCoord;
} // end "meanCoordFromMask()" -----------------------------------------------------------------
// Function to calculate the mean value of
// a set of n vectors each of dimension n
// namely a (n x n) matrix
vector<double> CNelderMead::VMean(vector<vector<double> > X, int n) {
vector<double> meanX(n);
for (int i=0; i<=n-1; i++) {
meanX[i]=0.0;
for (int j=0; j<=n-1; j++)
meanX[i] += X[i][j];
meanX[i] = meanX[i] / n;
}
return meanX;
}
bool
iAIDA::AIDA_Histogram_native::AIDA_Histogram2D::setRms( double rmsX, double rmsY )
{
const double sw = sumBinHeights();
const double mnX = meanX();
const double mnY = meanY();
m_sumWeightTimesSquaredX = ( mnX*mnX + rmsX*rmsX) * sw;
m_sumWeightTimesSquaredY = ( mnY*mnY + rmsY*rmsY) * sw;
m_validStatistics = false;
return true;
}
void
iAIDA::AIDA_Histogram_native::AIDA_Histogram2D::updateAnnotation() const
{
iAIDA::AIDA_Histogram_native::AIDA_BaseHistogram::updateAnnotation();
const AIDA::IAnnotation& anno = annotationNoUpdate();
AIDA::IAnnotation& annotation = const_cast< AIDA::IAnnotation& >( anno );
annotation.setValue( meanXKey, iAIDA_annotationNumberFormater.formatDouble( meanX() ) );
annotation.setValue( rmsXKey, iAIDA_annotationNumberFormater.formatDouble( rmsX() ) );
annotation.setValue( meanYKey, iAIDA_annotationNumberFormater.formatDouble( meanY() ) );
annotation.setValue( rmsYKey, iAIDA_annotationNumberFormater.formatDouble( rmsY() ) );
annotation.setValue( extra_entriesKey, iAIDA_annotationNumberFormater.formatInteger( extraEntries() ) );
}
static LogisticRegression _Table_to_LogisticRegression (Table me, long *factors, long numberOfFactors, long dependent1, long dependent2) {
long numberOfParameters = numberOfFactors + 1;
long numberOfCells = my rows -> size, numberOfY0 = 0, numberOfY1 = 0, numberOfData = 0;
double logLikelihood = 1e300, previousLogLikelihood = 2e300;
if (numberOfParameters < 1) // includes intercept
Melder_throw ("Not enough columns (has to be more than 1).");
/*
* Divide up the contents of the table into a number of independent variables (x) and two dependent variables (y0 and y1).
*/
autoNUMmatrix <double> x (1, numberOfCells, 0, numberOfFactors); // column 0 is the intercept
autoNUMvector <double> y0 (1, numberOfCells);
autoNUMvector <double> y1 (1, numberOfCells);
autoNUMvector <double> meanX (1, numberOfFactors);
autoNUMvector <double> stdevX (1, numberOfFactors);
autoNUMmatrix <double> smallMatrix (0, numberOfFactors, 0, numberOfParameters);
autoLogisticRegression thee = LogisticRegression_create (my columnHeaders [dependent1]. label, my columnHeaders [dependent2]. label);
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
double minimum = Table_getMinimum (me, factors [ivar]);
double maximum = Table_getMaximum (me, factors [ivar]);
Regression_addParameter (thee.peek(), my columnHeaders [factors [ivar]]. label, minimum, maximum, 0.0);
}
for (long icell = 1; icell <= numberOfCells; icell ++) {
y0 [icell] = Table_getNumericValue_Assert (me, icell, dependent1);
y1 [icell] = Table_getNumericValue_Assert (me, icell, dependent2);
numberOfY0 += y0 [icell];
numberOfY1 += y1 [icell];
numberOfData += y0 [icell] + y1 [icell];
x [icell] [0] = 1.0; /* Intercept. */
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
x [icell] [ivar] = Table_getNumericValue_Assert (me, icell, factors [ivar]);
meanX [ivar] += x [icell] [ivar] * (y0 [icell] + y1 [icell]);
}
}
if (numberOfY0 == 0 && numberOfY1 == 0)
Melder_throw ("No data in either class. Cannot determine result.");
if (numberOfY0 == 0)
Melder_throw ("No data in class ", my columnHeaders [dependent1]. label, ". Cannot determine result.");
if (numberOfY1 == 0)
Melder_throw ("No data in class ", my columnHeaders [dependent2]. label, ". Cannot determine result.");
/*
* Normalize the data.
*/
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
meanX [ivar] /= numberOfData;
for (long icell = 1; icell <= numberOfCells; icell ++) {
x [icell] [ivar] -= meanX [ivar];
}
}
for (long icell = 1; icell <= numberOfCells; icell ++) {
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
stdevX [ivar] += x [icell] [ivar] * x [icell] [ivar] * (y0 [icell] + y1 [icell]);
}
}
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
stdevX [ivar] = sqrt (stdevX [ivar] / numberOfData);
for (long icell = 1; icell <= numberOfCells; icell ++) {
x [icell] [ivar] /= stdevX [ivar];
}
}
/*
* Initial state of iteration: the null model.
*/
thy intercept = log ((double) numberOfY1 / (double) numberOfY0); // initial state of intercept: best guess for average log odds
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
RegressionParameter parm = static_cast<RegressionParameter> (thy parameters -> item [ivar]);
parm -> value = 0.0; // initial state of dependence: none
}
long iteration = 1;
for (; iteration <= 100; iteration ++) {
previousLogLikelihood = logLikelihood;
for (long ivar = 0; ivar <= numberOfFactors; ivar ++) {
for (long jvar = ivar; jvar <= numberOfParameters; jvar ++) {
smallMatrix [ivar] [jvar] = 0.0;
}
}
/*
* Compute the current log likelihood.
*/
logLikelihood = 0.0;
for (long icell = 1; icell <= numberOfCells; icell ++) {
double fittedLogit = thy intercept, fittedP, fittedQ, fittedLogP, fittedLogQ, fittedPQ, fittedVariance;
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
RegressionParameter parm = static_cast<RegressionParameter> (thy parameters -> item [ivar]);
fittedLogit += parm -> value * x [icell] [ivar];
}
/*
* Basically we have fittedP = 1.0 / (1.0 + exp (- fittedLogit)),
* but that works neither for fittedP values near 0 nor for values near 1.
*/
if (fittedLogit > 15.0) {
/*
* For large fittedLogit, fittedLogP = ln (1/(1+exp(-fittedLogit))) = -ln (1+exp(-fittedLogit)) =~ - exp(-fittedLogit)
*/
fittedLogP = - exp (- fittedLogit);
fittedLogQ = - fittedLogit;
fittedPQ = exp (- fittedLogit);
fittedP = exp (fittedLogP);
fittedQ = 1.0 - fittedP;
} else if (fittedLogit < -15.0) {
fittedLogP = fittedLogit;
fittedLogQ = - exp (fittedLogit);
fittedPQ = exp (fittedLogit);
fittedP = exp (fittedLogP);
fittedQ = 1 - fittedP;
} else {
fittedP = 1.0 / (1.0 + exp (- fittedLogit));
fittedLogP = log (fittedP);
fittedQ = 1.0 - fittedP;
fittedLogQ = log (fittedQ);
fittedPQ = fittedP * fittedQ;
}
logLikelihood += -2 * (y1 [icell] * fittedLogP + y0 [icell] * fittedLogQ);
/*
* Matrix shifting stuff.
* Suppose a + b Sk + c Tk = ln (pk / qk),
* where {a, b, c} are the coefficients to be optimized,
* Sk and Tk are properties of stimulus k,
* and pk and qk are the fitted probabilities for y1 and y0, respectively, given stimulus k.
* Then ln pk = - ln (1 + qk / pk) = - ln (1 + exp (- (a + b Sk + c Tk)))
* d ln pk / da = 1 / (1 + exp (a + b Sk + c Tk)) = qk
* d ln pk / db = qk Sk
* d ln pk / dc = qk Tk
* d ln qk / da = - pk
* Now LL = Sum(k) (y1k ln pk + y0k ln qk)
* so that dLL/da = Sum(k) (y1k d ln pk / da + y0k ln qk / da) = Sum(k) (y1k qk - y0k pk)
*/
fittedVariance = fittedPQ * (y0 [icell] + y1 [icell]);
for (long ivar = 0; ivar <= numberOfFactors; ivar ++) {
/*
* The last column gets the gradient of LL: dLL/da, dLL/db, dLL/dc.
*/
smallMatrix [ivar] [numberOfParameters] += x [icell] [ivar] * (y1 [icell] * fittedQ - y0 [icell] * fittedP);
for (long jvar = ivar; jvar <= numberOfFactors; jvar ++) {
smallMatrix [ivar] [jvar] += x [icell] [ivar] * x [icell] [jvar] * fittedVariance;
}
}
}
if (fabs (logLikelihood - previousLogLikelihood) < 1e-11) {
break;
}
/*
* Make matrix symmetric.
*/
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
for (long jvar = 0; jvar < ivar; jvar ++) {
smallMatrix [ivar] [jvar] = smallMatrix [jvar] [ivar];
}
}
/*
* Invert matrix in the simplest way, and shift and wipe the last column with it.
*/
for (long ivar = 0; ivar <= numberOfFactors; ivar ++) {
double pivot = smallMatrix [ivar] [ivar]; /* Save diagonal. */
smallMatrix [ivar] [ivar] = 1.0;
for (long jvar = 0; jvar <= numberOfParameters; jvar ++) {
smallMatrix [ivar] [jvar] /= pivot;
}
for (long jvar = 0; jvar <= numberOfFactors; jvar ++) {
if (jvar != ivar) {
double temp = smallMatrix [jvar] [ivar];
smallMatrix [jvar] [ivar] = 0.0;
for (long kvar = 0; kvar <= numberOfParameters; kvar ++) {
smallMatrix [jvar] [kvar] -= temp * smallMatrix [ivar] [kvar];
}
}
}
}
/*
* Update the parameters from the last column of smallMatrix.
*/
thy intercept += smallMatrix [0] [numberOfParameters];
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
RegressionParameter parm = static_cast<RegressionParameter> (thy parameters -> item [ivar]);
parm -> value += smallMatrix [ivar] [numberOfParameters];
}
}
if (iteration > 100) {
Melder_warning (L"Logistic regression has not converged in 100 iterations. The results are unreliable.");
}
for (long ivar = 1; ivar <= numberOfFactors; ivar ++) {
RegressionParameter parm = static_cast<RegressionParameter> (thy parameters -> item [ivar]);
parm -> value /= stdevX [ivar];
thy intercept -= parm -> value * meanX [ivar];
}
return thee.transfer();
}
void EdgeBoxGenerator::clusterEdges( arrayf &E, arrayf &O, arrayf &V )
{
int c, r, cd, rd, i, j; h=E._h; w=E._w;
// greedily merge connected edge pixels into clusters (create _segIds)
_segIds.init(h,w); _segCnt=1;
for( c=0; c<w; c++ ) for( r=0; r<h; r++ ) {
if( c==0 || r==0 || c==w-1 || r==h-1 || E.val(c,r)<=_edgeMinMag )
_segIds.val(c,r)=-1; else _segIds.val(c,r)=0;
}
for( c=1; c<w-1; c++ ) for( r=1; r<h-1; r++ ) {
if(_segIds.val(c,r)!=0) continue;
float sumv=0; int c0=c, r0=r; vectorf vs; vectori cs, rs;
while( sumv < _edgeMergeThr ) {
_segIds.val(c0,r0)=_segCnt;
float o0 = O.val(c0,r0), o1, v; bool found;
for( cd=-1; cd<=1; cd++ ) for( rd=-1; rd<=1; rd++ ) {
if( _segIds.val(c0+cd,r0+rd)!=0 ) continue; found=false;
for( i=0; i<cs.size(); i++ )
if( cs[i]==c0+cd && rs[i]==r0+rd ) { found=true; break; }
if( found ) continue; o1=O.val(c0+cd,r0+rd);
v=fabs(o1-o0)/PI; if(v>.5) v=1-v;
vs.push_back(v); cs.push_back(c0+cd); rs.push_back(r0+rd);
}
float minv=1000; j=0;
for( i=0; i<vs.size(); i++ ) if( vs[i]<minv ) {
minv=vs[i]; c0=cs[i]; r0=rs[i]; j=i;
}
sumv+=minv; if(minv<1000) vs[j]=1000;
}
_segCnt++;
}
// merge or remove small segments
_segMag.resize(_segCnt,0);
for( c=1; c<w-1; c++ ) for( r=1; r<h-1; r++ )
if( (j=_segIds.val(c,r))>0 ) _segMag[j]+=E.val(c,r);
for( c=1; c<w-1; c++ ) for( r=1; r<h-1; r++ )
if( (j=_segIds.val(c,r))>0 && _segMag[j]<=_clusterMinMag)
_segIds.val(c,r)=0;
i=1; while(i>0) {
i=0; for( c=1; c<w-1; c++ ) for( r=1; r<h-1; r++ ) {
if( _segIds.val(c,r)!=0 ) continue;
float o0=O.val(c,r), o1, v, minv=1000; j=0;
for( cd=-1; cd<=1; cd++ ) for( rd=-1; rd<=1; rd++ ) {
if( _segIds.val(c+cd,r+rd)<=0 ) continue; o1=O.val(c+cd,r+rd);
v=fabs(o1-o0)/PI; if(v>.5) v=1-v;
if( v<minv ) { minv=v; j=_segIds.val(c+cd,r+rd); }
}
_segIds.val(c,r)=j; if(j>0) i++;
}
}
// compactify representation
_segMag.assign(_segCnt,0); vectori map(_segCnt,0); _segCnt=1;
for( c=1; c<w-1; c++ ) for( r=1; r<h-1; r++ )
if( (j=_segIds.val(c,r))>0 ) _segMag[j]+=E.val(c,r);
for( i=0; i<_segMag.size(); i++ ) if( _segMag[i]>0 ) map[i]=_segCnt++;
for( c=1; c<w-1; c++ ) for( r=1; r<h-1; r++ )
if( (j=_segIds.val(c,r))>0 ) _segIds.val(c,r)=map[j];
// compute positional means and recompute _segMag
_segMag.assign(_segCnt,0); vectorf meanX(_segCnt,0), meanY(_segCnt,0);
vectorf meanOx(_segCnt,0), meanOy(_segCnt,0), meanO(_segCnt,0);
for( c=1; c<w-1; c++ ) for( r=1; r<h-1; r++ ) {
j=_segIds.val(c,r); if(j<=0) continue;
float m=E.val(c,r), o=O.val(c,r); _segMag[j]+=m;
meanOx[j]+=m*cos(2*o); meanOy[j]+=m*sin(2*o);
meanX[j]+=m*c; meanY[j]+=m*r;
}
for( i=0; i<_segCnt; i++ ) if( _segMag[i]>0 ) {
float m=_segMag[i]; meanX[i]/=m; meanY[i]/=m;
meanO[i]=atan2(meanOy[i]/m,meanOx[i]/m)/2;
}
// compute segment affinities
_segAff.resize(_segCnt); _segAffIdx.resize(_segCnt);
for(i=0; i<_segCnt; i++) _segAff[i].resize(0);
for(i=0; i<_segCnt; i++) _segAffIdx[i].resize(0);
const int rad = 2;
for( c=rad; c<w-rad; c++ ) for( r=rad; r<h-rad; r++ ) {
int s0=_segIds.val(c,r); if( s0<=0 ) continue;
for( cd=-rad; cd<=rad; cd++ ) for( rd=-rad; rd<=rad; rd++ ) {
int s1=_segIds.val(c+cd,r+rd); if(s1<=s0) continue;
bool found = false; for(i=0;i<_segAffIdx[s0].size();i++)
if(_segAffIdx[s0][i] == s1) { found=true; break; }
if( found ) continue;
float o=atan2(meanY[s0]-meanY[s1],meanX[s0]-meanX[s1])+PI/2;
float a=fabs(cos(meanO[s0]-o)*cos(meanO[s1]-o)); a=pow(a,_gamma);
_segAff[s0].push_back(a); _segAffIdx[s0].push_back(s1);
_segAff[s1].push_back(a); _segAffIdx[s1].push_back(s0);
}
}
// compute _segC and _segR
_segC.resize(_segCnt); _segR.resize(_segCnt);
for( c=1; c<w-1; c++ ) for( r=1; r<h-1; r++ )
if( (j=_segIds.val(c,r))>0 ) { _segC[j]=c; _segR[j]=r; }
// optionally create visualization (assume memory initialized is 3*w*h)
if( V._x ) for( c=0; c<w; c++ ) for( r=0; r<h; r++ ) {
i=_segIds.val(c,r);
V.val(c+w*0,r) = i<=0 ? 1 : ((123*i + 128)%255)/255.0f;
V.val(c+w*1,r) = i<=0 ? 1 : ((7*i + 3)%255)/255.0f;
V.val(c+w*2,r) = i<=0 ? 1 : ((174*i + 80)%255)/255.0f;
}
}
