/** Flag = 0: marginloss, no label loss. The label loss will always be zero
1: marginloss, and label loss.
*/
void CSMMMulticlassLoss::ComputeLoss(vector<unsigned int> y, vector<unsigned int> ylabel, vector<unsigned int> ybar, vector<unsigned int> ybarlabel, const CSeqMulticlassFeature::seqfeature_struct &x, const TheMatrix &w, double & marginloss, double & labelloss, int flag)
{
unsigned int i;
double w_dot_phi1 = 0;
double w_dot_phi2 = 0;
marginloss = 0;
unsigned int start;
if(is_first_phi1_used)
start = 0;
else
start = 1;
for(i=start; i < ybar.size(); i++)
{
_data->TensorPhi1(x.phi_1[ybar[i]],ybarlabel[i],0,tphi_1);
//tphi_1->Print();
w.Dot(*(tphi_1), w_dot_phi1);
marginloss += w_dot_phi1;
//printf("%d(%d):%2.4f\t",ybar[i],ybarlabel[i],marginloss);
}
for(i=1;i<ybar.size();i++)
{
int vb = 0;
_data->TensorPhi2(x.phi_2[ybar[i-1]][ybar[i]-ybar[i-1]-1], ybarlabel[i-1], ybarlabel[i], 0,vb,tphi_2);
w.Dot(*(tphi_2), w_dot_phi2);
marginloss += w_dot_phi2;
}
if(ybar.size() > 0)
{
//grad.Add(*(X[i].phi_2[ybar[ybar.size()-1]][X[i].len-1 - ybar[ybar.size()-1]-1]));////
_data->TensorPhi2(x.phi_2[ybar[ybar.size()-1]][x.len - ybar[ybar.size()-1]-1 ], ybarlabel[ybar.size()-1], 0, 0,0,tphi_2);
w.Dot(*(tphi_2), w_dot_phi2);
marginloss += w_dot_phi2;
}
//vector <unsigned int> yss = Boundry2StatSequence(y,ylabel,x.len);
//vector <unsigned int> ybarss = Boundry2StatSequence(ybar,ybarlabel,x.len);
//labelloss = Labelloss(yss,ybarss);
labelloss = AllDelta(ybar,y,ybarlabel,ylabel,x.len);
}
/** find best label (without label loss): g(w) := max_y' <w,\phi(x,y')>
*
* @param x [read] sequence
* @param w [read] weight vector
* @param ybar [write] found best label
* @param marginloss [write] margin loss <w,\Phi(x,y')> w.r.t to best y'
*/
void CSMMMulticlassLoss::find_best_label_grammer(const CSeqMulticlassFeature::seqfeature_struct &x, const TheMatrix &w, vector<unsigned int> &ybar, vector<unsigned int> &ybarlabel, double &marginloss, unsigned int personid, unsigned int classNum)
{
using namespace std;
// reset return values
marginloss = 0;
ybar.clear();
ybarlabel.clear();
/** The margin value vector used in dynamic programming
*/
vector< vector<double> > M (x.len+1,vector<double> (classNum,0));
/** The back pointers vector used in dynamic programming to retrieve the optimal path
*/
// The positions
vector< vector<int> > A (x.len+1,vector<int> (classNum,-1));
// The class labels
vector< vector<int> > C (x.len+1,vector<int> (classNum,0));
double maxval = -SML::INFTY;
double w_dot_phi1 = 0;
double w_dot_phi2 = 0;
double marginval = 0;
unsigned int right = 0;
unsigned int left = 0;
unsigned int start = 0;
unsigned int end = 0;
unsigned int classID = 0;
unsigned int classIDPrev = 0;
double sum = 0;
// compute DP statistics for positions 1 to len-1
for(classID=0;classID<classNum;classID++)
{
A[1][classID] = 0;
//C[1][classID] = 0;
}
if(is_first_phi1_used)
{
right =0;
for(classID=0;classID<classNum;classID++)
{
maxval = -SML::INFTY;
w_dot_phi1 = 0.0;
_data->TensorPhi1(x.phi_1[right],classID,0,tphi_1);
//tphi_1->Print();
w.Dot(*(tphi_1), w_dot_phi1);
marginval = w_dot_phi1;
sum = marginval;
if(sum > maxval)
{
M[right][classID] = marginval;
maxval = sum;
}
}
}
for(right=1; right < x.len+1; right++)
{
for(classID=0;classID<classNum;classID++)
{
// \Phi = (phi1, phi2[left,right])
// <w, \Phi> = <w,phi1> + <w,phi[left,right]>
maxval = -SML::INFTY;
w_dot_phi1 = 0.0;
if(right<x.len)
{
_data->TensorPhi1(x.phi_1[right],classID,0,tphi_1);
w.Dot(*(tphi_1), w_dot_phi1);
}
start = max(0,int(right-maxDuration));
//end = right;//-minDuration+1;
if(lastDuration>0)
{
unsigned int lastpos = x.len-lastDuration+1 ;
end = MIN(right,lastpos);
}
else
end = right;
for(left=start; left < end; left++)
{
classIDPrev = classID;
int vb = 0;
_data->TensorPhi2(x.phi_2[left][right-left-1], classIDPrev, classID, 0,vb,tphi_2);
w.Dot(*(tphi_2), w_dot_phi2);
marginval = w_dot_phi1 + w_dot_phi2;
sum = M[left][classIDPrev]+marginval;
if(sum > maxval)
{
A[right][classID] = left;
C[right][classID] = classIDPrev;
M[right][classID] = M[left][classIDPrev] + marginval;
maxval = sum;
}
}
}
}
// get optimal path (i.e. segmentation)
unsigned int pos,prepos,classid,preclassid;
int maxclassid = 0;
maxval = -SML::INFTY;
for(unsigned int i=0;i<classNum;i++)
{
sum = M[x.len][i];
if(sum>maxval)
{
maxval = sum;
maxclassid = i;
}
}
pos = A[x.len][maxclassid];
classid = C[x.len][maxclassid];
if(lastDuration>0)
{
pos = x.len-lastDuration;
classid = 0;
}
ybar.push_back(pos);
ybarlabel.push_back(classid);
prepos = pos;
preclassid = classid;
while(A[pos][classid] >= 0)
{
pos = A[prepos][preclassid];
classid = C[prepos][preclassid];
ybar.push_back(pos);//positions
ybarlabel.push_back(classid);//class labels
prepos = pos;
preclassid = classid;
}
marginloss = M[x.len][maxclassid];
reverse(ybar.begin(), ybar.end());
reverse(ybarlabel.begin(), ybarlabel.end());
}
/** find best label with a grammer(with label loss): g(w) := max_y' <w,\phi(x,y')> + Delta(y', y)
*
* @param x [read] sequence
* @param y [read] actual label for x
* @param w [read] weight vector
* @param ybar [write] found best label
* @param marginloss [write] margin loss <w,\Phi(x,y')> w.r.t to best y'
* @param labelloss [write] label loss \Delta(y',y) w.r.t. to best y'
*
*/
void CSMMMulticlassLoss::find_best_label_grammer(const vector<unsigned int> &y,const vector<unsigned int> &ylabel, const CSeqMulticlassFeature::seqfeature_struct &x, const TheMatrix &w, vector<unsigned int> &ybar,vector<unsigned int> &ybarlabel, double &marginloss, double &labelloss, unsigned int personid, unsigned int classNum)
{
// reset return values
marginloss = 0;
labelloss = 0;
ybar.clear();
ybarlabel.clear();
/** The margin value vector used in dynamic programming
*/
vector< vector<double> > M (x.len+1,vector<double> (classNum,0));
/** The label loss value vector used in dynamic programming
*/
vector< vector<double> > L (x.len+1,vector<double> (classNum,0));
/** The back pointers vector used in dynamic programming to retrieve the optimal path
*/
// The positions
vector< vector<int> > A (x.len+1,vector<int> (classNum,-1));
// The class labels
vector< vector<int> > C (x.len+1,vector<int> (classNum,0));
double maxval = -SML::INFTY;
double w_dot_phi1 = 0;
double w_dot_phi2 = 0;
double marginval = 0;
double labelval = 0;
unsigned int right = 0;
unsigned int left = 0;
unsigned int start = 0;
unsigned int end = 0;
unsigned int classID = 0;
unsigned int classIDPrev = 0;
double sum = 0;
// compute DP statistics for positions 1 to len-1
// L[0] += y.size()-2;
// A[1] = 0;
for(classID=0;classID<classNum;classID++)
{
A[1][classID] = 0;
//C[1][classID] = 0;
}
//debug
//printf("x.len:%d",x.len);
if(is_first_phi1_used)
{
right =0;
for(classID=0;classID<classNum;classID++)
{
maxval = -SML::INFTY;
w_dot_phi1 = 0.0;
_data->TensorPhi1(x.phi_1[right],classID,0,tphi_1);
//tphi_1->Print();
w.Dot(*(tphi_1), w_dot_phi1);
marginval = w_dot_phi1;
sum = marginval;
if(sum > maxval)
{
M[right][classID] = marginval;
maxval = sum;
}
}
}
for(right=1; right < x.len+1; right++)
{
for(classID=0;classID<classNum;classID++)
{
// \Phi = (phi1, phi2[left,right])
// <w, \Phi> = <w,phi1> + <w,phi[left,right]>
maxval = -SML::INFTY;
w_dot_phi1 = 0.0;
//w.Dot(*(x.phi_1[right]), w_dot_phi1);
//printf("pos:%d,classid:%d ",right,classID);fflush(stdout);
//x.phi_1[right]->Print();
if(right<x.len)
{
_data->TensorPhi1(x.phi_1[right],classID,0,tphi_1);
//tphi_1->Print();
w.Dot(*(tphi_1), w_dot_phi1);
}
start = max(0,int(right-maxDuration));
//end = right;//-minDuration+1;
if(lastDuration>0)
{
unsigned int lastpos = x.len-lastDuration+1 ;
end = MIN(right,lastpos);
}
else
end = right;
for(left=start; left < end; left++)
{
classIDPrev = classID;
labelval = PartialDelta(left,right,y,ylabel,classIDPrev,x.len);
assert( (labelval<=x.len) && (labelval>=0) );
int vb = 0;
_data->TensorPhi2(x.phi_2[left][right-left-1], classIDPrev, classID, 0,vb,tphi_2);
w.Dot(*(tphi_2), w_dot_phi2);
marginval = w_dot_phi1 + w_dot_phi2;
sum = M[left][classIDPrev]+marginval + L[left][classIDPrev]+labelval;
if(sum > maxval)
{
A[right][classID] = left;
C[right][classID] = classIDPrev;
M[right][classID] = M[left][classIDPrev] + marginval;
L[right][classID] = L[left][classIDPrev] + labelval;
maxval = sum;
}
}
}
}
// get optimal path (i.e. segmentation)
unsigned int pos,prepos,classid,preclassid;
int maxclassid = 0;
maxval = -SML::INFTY;
for(unsigned int i=0;i<classNum;i++)
{
sum = M[x.len][i] + L[x.len][i];
if(sum>maxval)
{
maxval = sum;
maxclassid = i;
}
}
pos = A[x.len][maxclassid];
classid = C[x.len][maxclassid];
if(lastDuration>0)
{
pos = x.len-lastDuration;
classid = 0;
}
ybar.push_back(pos);
ybarlabel.push_back(classid);
prepos = pos;
preclassid = classid;
while(A[pos][classid] >= 0)
{
pos = A[prepos][preclassid];
classid = C[prepos][preclassid];
ybar.push_back(pos);//positions
ybarlabel.push_back(classid);//class labels
//printf("%d(%d):%2.4f ",pos,classid,L[pos][classid]);fflush(stdout);
prepos = pos;
preclassid = classid;
}
marginloss = M[x.len][maxclassid];
labelloss = L[x.len][maxclassid];
//printf("finished back track\n labelloss:%3.4f,marginloss:%3.4f\n",labelloss,marginloss);fflush(stdout);
reverse(ybar.begin(), ybar.end());
reverse(ybarlabel.begin(), ybarlabel.end());
//printf("reversed\n");fflush(stdout);
unsigned int i;
if(verbosity>=2)
{
printf("y: ");
for(i=0;i<y.size();i++)
{
printf("%d(%d) ",y[i],ylabel[i]);
}
fflush(stdout);
printf("\nybar:");
for(i=0;i<ybar.size();i++)
{
printf("%d(%d) ",ybar[i],ybarlabel[i]);
}
fflush(stdout);
printf("\nmargin:%f, loss:%f, totalloss:%f\n",marginloss,labelloss,marginloss+labelloss);
}
}
