double
InvLogitGaussianJointPdf<V,M>::lnValue(
const V& domainVector,
const V* domainDirection,
V* gradVector,
M* hessianMatrix,
V* hessianEffect) const
{
double returnValue;
double lnDeterminant = 0.0;
V transformedDomainVector(domainVector);
V min_domain_bounds(this->m_domainBoxSubset.minValues());
V max_domain_bounds(this->m_domainBoxSubset.maxValues());
double lnjacobian = 0.0;
for (unsigned int i = 0; i < domainVector.sizeLocal(); i++) {
double min_val = min_domain_bounds[i];
double max_val = max_domain_bounds[i];
if (queso_isfinite(min_val) &&
queso_isfinite(max_val)) {
if (domainVector[i] == min_val || domainVector[i] == max_val) {
// Exit early if we can
return -INFINITY;
}
// Left- and right-hand sides are finite. Do full transform.
transformedDomainVector[i] = std::log(domainVector[i] - min_val) -
std::log(max_val - domainVector[i]);
lnjacobian += std::log(max_val - min_val) -
std::log(domainVector[i] - min_val) -
std::log(max_val - domainVector[i]);
}
else if (queso_isfinite(min_val) &&
!queso_isfinite(max_val)) {
if (domainVector[i] == min_val) {
// Exit early if we can
return -INFINITY;
}
// Left-hand side finite, but right-hand side is not.
// Do only left-hand transform.
transformedDomainVector[i] = std::log(domainVector[i] - min_val);
lnjacobian += -std::log(domainVector[i] - min_val);
}
else if (!queso_isfinite(min_val) &&
queso_isfinite(max_val)) {
if (domainVector[i] == max_val) {
// Exit early if we can
return -INFINITY;
}
// Right-hand side is finite, but left-hand side is not.
// Do only right-hand transform.
transformedDomainVector[i] = -std::log(max_val - domainVector[i]);
lnjacobian += -std::log(max_val - domainVector[i]);
}
else {
// No transform.
transformedDomainVector[i] = domainVector[i];
}
}
V diffVec(transformedDomainVector - this->lawExpVector());
if (m_diagonalCovMatrix) {
returnValue = ((diffVec * diffVec) /
this->lawVarVector()).sumOfComponents();
if (m_normalizationStyle == 0) {
unsigned int iMax = this->lawVarVector().sizeLocal();
for (unsigned int i = 0; i < iMax; ++i) {
lnDeterminant += log(this->lawVarVector()[i]);
}
}
}
else {
V tmpVec = this->m_lawCovMatrix->invertMultiply(diffVec);
returnValue = (diffVec * tmpVec).sumOfComponents();
if (m_normalizationStyle == 0) {
lnDeterminant = this->m_lawCovMatrix->lnDeterminant();
}
}
if (m_normalizationStyle == 0) {
returnValue += ((double) this->lawVarVector().sizeLocal()) * log(2 * M_PI);
returnValue += lnDeterminant;
}
returnValue *= -0.5;
returnValue += m_logOfNormalizationFactor;
returnValue += lnjacobian;
return returnValue;
}
double
GaussianJointPdf<V,M>::lnValue(
const V& domainVector,
const V* domainDirection,
V* gradVector,
M* hessianMatrix,
V* hessianEffect) const
{
if ((m_env.subDisplayFile()) && (m_env.displayVerbosity() >= 55)) {
*m_env.subDisplayFile() << "Entering GaussianJointPdf<V,M>::lnValue()"
<< ", meanVector = " << *m_lawExpVector
<< ", lawCovMatrix = " << *m_lawCovMatrix
<< ": domainVector = " << domainVector
<< std::endl;
}
queso_require_msg(!(domainDirection || hessianMatrix || hessianEffect), "incomplete code for gradVector, hessianMatrix and hessianEffect calculations");
double returnValue = 0.;
double lnDeterminant = 0.;
if (this->m_domainSet.contains(domainVector) == false) {
// What should the gradient be here?
returnValue = -INFINITY;
}
else {
V diffVec(domainVector - this->lawExpVector());
if (m_diagonalCovMatrix) {
returnValue = ((diffVec*diffVec)/this->lawVarVector()).sumOfComponents();
// Compute the gradient of log of the pdf.
// The log of a Gaussian pdf is:
// f(x) = - 1/2 (x - \mu)^T \Sigma^{-1} (x - \mu)
// Therefore
// \frac{df}{dx}(x) = - (x - \mu)^T \Sigma^{-1}
// = - (\Sigma^{-1}^T (x - \mu))^T
// = - (\Sigma^{-1} (x - \mu))^T
// = - \Sigma^{-1} (x - \mu) (row/column vector doesn't matter)
//
// So if \Sigma is diagonal we have a component-wise product of two
// vectors (x - \mu) and the diagonal elements of \Sigma^{-1}
if (gradVector) {
(*gradVector) = diffVec; // Copy
(*gradVector) /= this->lawVarVector();
(*gradVector) *= -1.0;
}
if (m_normalizationStyle == 0) {
unsigned int iMax = this->lawVarVector().sizeLocal();
for (unsigned int i = 0; i < iMax; ++i) {
lnDeterminant += std::log(this->lawVarVector()[i]);
}
}
}
else {
V tmpVec = this->m_lawCovMatrix->invertMultiply(diffVec);
returnValue = (diffVec*tmpVec).sumOfComponents();
// Compute the gradient of log of the pdf.
// The log of a Gaussian pdf is:
// f(x) = - 1/2 (x - \mu)^T \Sigma^{-1} (x - \mu)
// Therefore
// \frac{df}{dx}(x) = - (x - \mu)^T \Sigma^{-1}
// = - (\Sigma^{-1}^T (x - \mu))^T
// = - (\Sigma^{-1} (x - \mu))^T
// = - \Sigma^{-1} (x - \mu) (row/column vector doesn't matter)
if (gradVector) {
(*gradVector) = tmpVec;
(*gradVector) *= -1.0;
}
if (m_normalizationStyle == 0) {
lnDeterminant = this->m_lawCovMatrix->lnDeterminant();
}
}
if (m_normalizationStyle == 0) {
returnValue += ((double) this->lawVarVector().sizeLocal()) * std::log(2*M_PI); // normalization of pdf
returnValue += lnDeterminant; // normalization of pdf
}
returnValue *= -0.5;
}
returnValue += m_logOfNormalizationFactor;
if ((m_env.subDisplayFile()) && (m_env.displayVerbosity() >= 99)) {
*m_env.subDisplayFile() << "Leaving GaussianJointPdf<V,M>::lnValue()"
<< ", m_normalizationStyle = " << m_normalizationStyle
<< ", m_diagonalCovMatrix = " << m_diagonalCovMatrix
<< ", m_logOfNormalizationFactor = " << m_logOfNormalizationFactor
<< ", lnDeterminant = " << lnDeterminant
<< ", meanVector = " << *m_lawExpVector
<< ", lawCovMatrix = " << *m_lawCovMatrix
<< ": domainVector = " << domainVector
<< ", returnValue = " << returnValue
<< std::endl;
}
return returnValue;
}
double
LogNormalJointPdf<V,M>::lnValue(
const V& domainVector,
const V* domainDirection,
V* gradVector,
M* hessianMatrix,
V* hessianEffect) const
{
if ((m_env.subDisplayFile()) && (m_env.displayVerbosity() >= 55)) {
*m_env.subDisplayFile() << "Entering LogNormalJointPdf<V,M>::lnValue()"
<< ", meanVector = " << *m_lawExpVector
<< ": domainVector = " << domainVector
<< std::endl;
}
UQ_FATAL_TEST_MACRO((gradVector || hessianMatrix || hessianEffect),
m_env.worldRank(),
"LogNormalJointPdf<V,M>::lnValue()",
"incomplete code for gradVector, hessianMatrix and hessianEffect calculations");
if (domainDirection) {}; // just to remove compiler warning
double returnValue = 0.;
V zeroVector(domainVector);
zeroVector.cwSet(0.);
if (domainVector.atLeastOneComponentSmallerOrEqualThan(zeroVector)) {
returnValue = -INFINITY;
}
else if (this->m_domainSet.contains(domainVector) == false) { // prudenci 2011-Oct-04
returnValue = -INFINITY;
}
else {
if (m_diagonalCovMatrix) {
V diffVec(zeroVector);
for (unsigned int i = 0; i < domainVector.sizeLocal(); ++i) {
diffVec[i] = std::log(domainVector[i]) - this->lawExpVector()[i];
}
returnValue = ((diffVec*diffVec)/this->lawVarVector()).sumOfComponents();
returnValue *= -0.5;
if (m_normalizationStyle == 0) {
for (unsigned int i = 0; i < domainVector.sizeLocal(); ++i) {
returnValue -= std::log(domainVector[i] * std::sqrt(2. * M_PI * this->lawVarVector()[i])); // Contribution of 1/(x\sqrt{2\pi\sigma^2})
}
}
}
else {
UQ_FATAL_TEST_MACRO(true,
m_env.worldRank(),
"LogNormalJointPdf<V,M>::lnValue()",
"situation with a non-diagonal covariance matrix makes no sense");
}
returnValue += m_logOfNormalizationFactor; // [PDF-10]
}
if ((m_env.subDisplayFile()) && (m_env.displayVerbosity() >= 55)) {
*m_env.subDisplayFile() << "Leaving LogNormalJointPdf<V,M>::lnValue()"
<< ", meanVector = " << *m_lawExpVector
<< ": domainVector = " << domainVector
<< ", returnValue = " << returnValue
<< std::endl;
}
return returnValue;
}
void MakeGraph( Energy<float,float,float> *e , int *vars , vector< set<int> > &neighbors, vector<int> &Disparity,
vector< vector<double> > Colors_L, vector<int> FLabels_L, vector<int> BLabels_L ,
vector<double> FDistVec_L , vector<double> BDistVec_L ,
vector< vector<double> > Colors_R, vector<int> FLabels_R, vector<int> BLabels_R ,
vector<double> FDistVec_R , vector<double> BDistVec_R ,
int numLabels , int num_high , int numRows , int numCols , double lambda_p , double lambda_c )
{
int i;
set<int>::iterator pIter;
const int K = 10000;
// Edge weight for infinity -- All other edge weights less than 1 -- so it suffices to have this weight as more than sum of all other edges, i.e. >8
double beta;
double countCand = 0;
double SumColors = 0;
for( i=0 ; i<numLabels ; i++ )
{
for ( pIter = neighbors[i].begin() ; pIter != neighbors[i].end() ; pIter++ )
{
SumColors += pow( diffVec(Colors_L[i],Colors_L[*pIter]) , 2 );
countCand++;
}
}
for( i=numLabels ; i<2*numLabels ; i++ )
{
for ( pIter = neighbors[i].begin() ; pIter != neighbors[i].end() ; pIter++ )
{
SumColors += pow( diffVec(Colors_R[i-numLabels],Colors_R[*pIter-numLabels]) , 2 );
countCand++;
}
}
beta = countCand/(2*SumColors);
///////////////////////////////////////////////////////////////////////////////////////////////////
//when reducing the order of high-order term, every high-order clique needs 5 new auxiliary variables
//calulate the factors for data term
vector<double> ForeEdges, BackEdges;
ForeEdges.clear();
BackEdges.clear();
printf("Graph Making code starts... \n");
/****** Making Terminal Edge Weights ******/
int count=0 ;
for( i=0; i<numLabels; i++)
{
if(findValVec(FLabels_L,i))
{
count++ ;
ForeEdges.push_back(K);
BackEdges.push_back(0);
}
else if(findValVec(BLabels_L,i))
{
count++ ;
ForeEdges.push_back(0);
BackEdges.push_back(K);
}
else
{
double temp;
temp = BDistVec_L[i]+FDistVec_L[i] ;
ForeEdges.push_back(FDistVec_L[i]/temp);
BackEdges.push_back(BDistVec_L[i]/temp);
}
}
for( i=0; i<numLabels; i++)
{
if(findValVec(FLabels_R,i))
{
count++ ;
ForeEdges.push_back(K);
BackEdges.push_back(0);
}
else if(findValVec(BLabels_R,i))
{
count++ ;
ForeEdges.push_back(0);
BackEdges.push_back(K);
}
else
{
double temp;
temp = BDistVec_R[i]+FDistVec_R[i] ;
ForeEdges.push_back(FDistVec_R[i]/temp);
BackEdges.push_back(BDistVec_R[i]/temp);
}
}
printf("count=%d \n Terminal Edge Weights Made... \n",count);
///////////////////////////////////////////////
/******* Start making the graph **************/
///////////////////////////////////////////////
// Add Nodes
for( i=0 ; i<2*numLabels+4*num_high*5 ; i++ )
vars[i] = e -> add_variable();
printf("Nodes Added to Graph... \n");
//data term
for( i=0 ; i<2*numLabels ; i++ )
{
e -> add_term1(vars[i], ForeEdges[i], BackEdges[i]);
}
printf("Terminal Edge Weights Set... \n");
//smooth term
for( i=0 ; i<numLabels ; i++ )
for (pIter = neighbors[i].begin(); pIter != neighbors[i].end(); pIter++)
{
int tmpN = *pIter;
double Energ = exp( -beta * pow(diffVec( Colors_L[i], Colors_L[tmpN]) ,2 ) );
e -> add_term2( vars[i], vars[tmpN] , 0 , lambda_p*Energ, lambda_p*Energ , 0);
}
for( i=numLabels ; i<2*numLabels ; i++ )
for (pIter = neighbors[i].begin(); pIter != neighbors[i].end(); pIter++)
{
int tmpN = *pIter;
double Energ = exp( -beta * pow(diffVec( Colors_R[i-numLabels], Colors_R[tmpN-numLabels]) ,2 ) );
e -> add_term2( vars[i], vars[tmpN] , 0 , lambda_p*Energ, lambda_p*Energ , 0);
}
//correspondence term: the low-order terms decomposed from the high-order term
int count_high =0 ;
int loc_y[4] = { -1 , 1, 0, 0 } ;
int loc_x[4] = { 0 , 0, -1, 1 } ; //up down left right
float f1 , f2 , a ;
f2 = 2 ;
float theta1 , theta2 , gamma1, gamma2 ;
theta1=0.5 ;
theta2=1.5 ; ////////////////////////////////
////////////////////////////////////////
gamma1 = 200 ;
gamma2 = 1000 ;
for( int c=1 ; c<numCols-1 ; c++ )
for( int h=1 ; h<numRows-1 ; h++ )
{
int this_ind , this_disp , cr ;
this_ind = c*numRows + h ;
this_disp = Disparity[this_ind] ;
cr = c- this_disp ;
if( cr >= 1 )
{
vector<double> Meanwin_l(3) ;
vector<double> Meanwin_r(3) ;
get_win_mean( Colors_L , h , c , numRows , numCols , Meanwin_l ) ;
get_win_mean( Colors_R , h , cr, numRows , numCols , Meanwin_r ) ;
double C_rl = exp( -sqrt( 1/(2*16*16) * pow(diffVec( Meanwin_l, Meanwin_r) ,2) ) );
C_rl *= lambda_c ;
float local_var=0;
local_var = get_var( Colors_L , Colors_R , h, c, cr , numRows , numCols , Meanwin_l , Meanwin_r);
if ( local_var<theta1 )
f1 = gamma1 ;
else if ( local_var>theta2 )
f1 = gamma2 ;
else
{
f1 = -(gamma2-gamma1)/(theta2-theta1)*local_var + gamma2*theta2-gamma1*theta1 ;
}
a = f2 - 2*f1 ;
if(a>0)
{
//one order terms transformed from the high-order term
e -> add_term1(vars[2*numLabels + count_high*5] , 0, -C_rl*2*a);
e -> add_term1(vars[2*numLabels + count_high*5 +1], 0, -C_rl*2*a);
e -> add_term1(vars[2*numLabels + count_high*5 +2], 0, -C_rl*2*a);
e -> add_term1(vars[2*numLabels + count_high*5 +3], 0, -C_rl*2*a);
e -> add_term1(vars[2*numLabels + count_high*5 +4], 0, -C_rl*12*a); // w1 w2 w3 w4 w
//two order terms transformed from the high-order term
for( int u=0 ; u<4 ; u++ )
{
//find the corresponding pixels and their neighbours
int neigh_y , neigh_xl , neigh_xr , ind_l , ind_r , neigh_ind_l , neigh_ind_r ;
neigh_y = h + loc_y[u] ;
neigh_xl = c + loc_x[u] ;
neigh_xr = cr + loc_x[u] ;
ind_l = c * numRows + h ;
ind_r = cr * numRows + h ;
neigh_ind_l = neigh_xl * numRows + neigh_y ;
neigh_ind_r = neigh_xr * numRows + neigh_y ;
e -> add_term1(vars[ind_l] , 0, C_rl*(f1+6*a) );
e -> add_term1(vars[ind_r+numLabels] , 0, C_rl*(f1+6*a) );
e -> add_term1(vars[neigh_ind_l] , 0, C_rl*(f1+6*a) );
e -> add_term1(vars[neigh_ind_r+numLabels] , 0, C_rl*(f1+6*a) ); //x0 y0 x1 y1
e -> add_term2( vars[ind_l] , vars[neigh_ind_l] , 0 , 0 , 0 , -C_rl*3*a); //x0x1 y0y1 x0y1 x1y0
e -> add_term2( vars[ind_r+numLabels], vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , -C_rl*3*a);
e -> add_term2( vars[ind_l] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , -C_rl*3*a);
e -> add_term2( vars[neigh_ind_l] , vars[ind_r+numLabels] , 0 , 0 , 0 , -C_rl*3*a);
e -> add_term2( vars[ind_l] , vars[ind_r+numLabels] , 0 , 0 , 0 , C_rl*(-2*f1-4*a) );
e -> add_term2( vars[neigh_ind_l], vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , C_rl*(-2*f1-4*a) ); //x0y0 x1y1
e -> add_term2( vars[2*numLabels + count_high*5] , vars[ind_l] , 0 , 0 , 0 , -C_rl*2*a ); //w1 *( x0 + x1 + y0 )
e -> add_term2( vars[2*numLabels + count_high*5] , vars[neigh_ind_l] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5] , vars[ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+1] , vars[ind_l] , 0 , 0 , 0 , -C_rl*2*a ); //w2 *( x0 + x1 + y1 )
e -> add_term2( vars[2*numLabels + count_high*5+1] , vars[neigh_ind_l] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+1] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+2] , vars[ind_l] , 0 , 0 , 0 , -C_rl*2*a ); //w3 *( x0 + y0 + y1 )
e -> add_term2( vars[2*numLabels + count_high*5+2] , vars[ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+2] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+3] , vars[neigh_ind_l] , 0 , 0 , 0 , -C_rl*2*a ); //w4 *( x1 + y0 + y1 )
e -> add_term2( vars[2*numLabels + count_high*5+3] , vars[ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+3] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+4] , vars[ind_l] , 0 , 0 , 0 , C_rl*4*a ); //w*(x0+x1+y0+y1)
e -> add_term2( vars[2*numLabels + count_high*5+4] , vars[neigh_ind_l] , 0 , 0 , 0 , C_rl*4*a );
e -> add_term2( vars[2*numLabels + count_high*5+4] , vars[ind_r+numLabels] , 0 , 0 , 0 , C_rl*4*a );
e -> add_term2( vars[2*numLabels + count_high*5+4] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , C_rl*4*a );
}
}
else
{
e -> add_term1(vars[2*numLabels + count_high*5] , 0, C_rl*4*a);
e -> add_term1(vars[2*numLabels + count_high*5 +1], 0, C_rl*4*a);
e -> add_term1(vars[2*numLabels + count_high*5 +2], 0, C_rl*4*a);
e -> add_term1(vars[2*numLabels + count_high*5 +3], 0, C_rl*4*a);
e -> add_term1(vars[2*numLabels + count_high*5 +4], 0, C_rl*12*a); // w1 w2 w3 w4 w
for( int u=0 ; u<4 ; u++ )
{
int neigh_y , neigh_xl , neigh_xr , ind_l , ind_r , neigh_ind_l , neigh_ind_r ;
neigh_y = h + loc_y[u] ;
neigh_xl = c + loc_x[u] ;
neigh_xr = cr + loc_x[u] ;
ind_l = c * numRows + h ;
ind_r = cr * numRows + h ;
neigh_ind_l = neigh_xl * numRows + neigh_y ;
neigh_ind_r = neigh_xr * numRows + neigh_y ;
e -> add_term1(vars[ind_l] , 0, C_rl*f1 );
e -> add_term1(vars[ind_r+numLabels] , 0, C_rl*f1 );
e -> add_term1(vars[neigh_ind_l] , 0, C_rl*f1 );
e -> add_term1(vars[neigh_ind_r+numLabels] , 0, C_rl*f1 ); //x0 y0 x1 y1
e -> add_term2( vars[ind_l] , vars[neigh_ind_l] , 0 , 0 , 0 , C_rl*5*a); //x0x1 y0y1 x0y1 x1y0
e -> add_term2( vars[ind_r+numLabels], vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , C_rl*5*a);
e -> add_term2( vars[ind_l] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , C_rl*5*a);
e -> add_term2( vars[neigh_ind_l] , vars[ind_r+numLabels] , 0 , 0 , 0 , C_rl*5*a);
e -> add_term2( vars[ind_l] , vars[ind_r+numLabels] , 0 , 0 , 0 , C_rl*(-2*f1+4*a) );
e -> add_term2( vars[neigh_ind_l], vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , C_rl*(-2*f1+4*a) ); //x0y0 x1y1
e -> add_term2( vars[2*numLabels + count_high*5] , vars[ind_l] , 0 , 0 , 0 , -C_rl*2*a ); //w1 *( x0 + x1 + y0 )
e -> add_term2( vars[2*numLabels + count_high*5] , vars[neigh_ind_l] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5] , vars[ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+1] , vars[ind_l] , 0 , 0 , 0 , -C_rl*2*a ); //w2 *( x0 + x1 + y1 )
e -> add_term2( vars[2*numLabels + count_high*5+1] , vars[neigh_ind_l] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+1] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+2] , vars[ind_l] , 0 , 0 , 0 , -C_rl*2*a ); //w3 *( x0 + y0 + y1 )
e -> add_term2( vars[2*numLabels + count_high*5+2] , vars[ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+2] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+3] , vars[neigh_ind_l] , 0 , 0 , 0 , -C_rl*2*a ); //w4 *( x1 + y0 + y1 )
e -> add_term2( vars[2*numLabels + count_high*5+3] , vars[ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+3] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , -C_rl*2*a );
e -> add_term2( vars[2*numLabels + count_high*5+4] , vars[ind_l] , 0 , 0 , 0 , -C_rl*8*a ); //w*(x0+x1+y0+y1)
e -> add_term2( vars[2*numLabels + count_high*5+4] , vars[neigh_ind_l] , 0 , 0 , 0 , -C_rl*8*a );
e -> add_term2( vars[2*numLabels + count_high*5+4] , vars[ind_r+numLabels] , 0 , 0 , 0 , -C_rl*8*a );
e -> add_term2( vars[2*numLabels + count_high*5+4] , vars[neigh_ind_r+numLabels] , 0 , 0 , 0 , -C_rl*8*a );
}
}
count_high ++ ;
}
}
printf("Graph Made... \n");
}