/**
* Compute loss and gradient of Huber hinge loss.
* CAUTION: f is passed by reference and is changed within this
* function. This is done for efficiency reasons, otherwise we would
* have had to create a new copy of f.
*
* @param loss [write] loss value computed.
* @param f [read/write] prediction vector.
* @param l [write] partial derivative of loss function w.r.t. f
*/
void CHuberHingeLoss::LossAndGrad(double& loss, TheMatrix& f, TheMatrix& l)
{
f.ElementWiseMult(_data->labels());
double* yf = f.Data();
double* Y = _data->labels().Data();
int len = f.Length();
loss = 0.0;
l.Zero();
for(int i=0; i < len; i++)
{
double v = 1-yf[i];
if(h < v)
{
loss += v;
l.Set(i,-Y[i]);
}
else if(-h > v) {}
else
{
loss += (v+h)*(v+h)/4/h;
l.Set(i, -Y[i]*(v+h)/2/h);
}
}
}
/**
* Compute loss and partial derivative of hinge loss w.r.t f
*
* @param loss [write] loss value computed.
* @param f [r/w] = X*w
* @param l [write] partial derivative of loss w.r.t. f
*/
void CLogisticLoss::LossAndGrad(double& loss, TheMatrix& f, TheMatrix& l)
{
l.Zero(); // for gradient computation i.e. grad := l'*X
f.ElementWiseMult(_data->labels());
double* f_array = f.Data(); // pointer to memory location of f (faster element access)
int len = f.Length();
double exp_yf = 0.0;
for(int i=0; i < len; i++)
{
if(fabs(f_array[i]) == 0.0)
{
loss += LN2;
l.Set(i,-0.5);
}
else if (f_array[i] > 0.0)
{
exp_yf = exp(-f_array[i]);
loss += log(1+exp_yf);
l.Set(i,-exp_yf/(1+exp_yf));
}
else
{
exp_yf = exp(f_array[i]);
loss += log(1+exp_yf) - f_array[i];
l.Set(i,-1.0/(1+exp_yf));
}
}
l.ElementWiseMult(_data->labels());
}
/** The subgradient is chosen as sgn(w)
*/
void CL1N1::ComputeRegAndGradient(CModel& model, double& reg, TheMatrix& grad)
{
reg = 0;
TheMatrix &w = model.GetW();
w.Norm1(reg);
grad.Zero();
for(int i=0; i<w.Length(); i++)
{
double val = 0;
w.Get(i,val);
grad.Set(i,SML::sgn(val));
}
}
/**
* Compute loss and gradient of Least Absolute Deviation loss w.r.t f
*
* @param loss [write] loss value computed.
* @param f [r/w] = X*w
* @param l [write] partial derivative of loss w.r.t. f
*/
void CLeastAbsDevLoss::LossAndGrad(double& loss, TheMatrix& f, TheMatrix& l)
{
loss = 0;
l.Zero();
double *Y_array = _data->labels().Data();
double* f_array = f.Data();
int len = f.Length();
for(int i=0; i < len; i++)
{
double f_minus_y = f_array[i] - Y_array[i];
loss += fabs(f_minus_y);
l.Set(i, SML::sgn(f_minus_y));
}
}
/**
* Compute loss and gradient of novelty detection loss.
* CAUTION: f is passed by reference and is changed within this
* function. This is done for efficiency reasons, otherwise we would
* have had to create a new copy of f.
*
* @param loss [write] loss value computed.
* @param f [read/write] prediction vector.
* @param l [write] partial derivative of loss function w.r.t. f
*/
void CNoveltyLoss::LossAndGrad(double& loss, TheMatrix& f, TheMatrix& l)
{
double* f_array = f.Data(); // pointer to memory location of f (faster element access)
int len = f.Length();
l.Zero(); // grad := l'*X
for(int i=0; i < len; i++)
{
if(rho > f_array[i])
{
loss += rho - f_array[i];
l.Set(i, -1.0);
}
}
}
/**
* Compute loss and partial derivative of NDCGRank loss w.r.t f
*
* @param loss [write] loss value computed.
* @param f [r/w] = X*w
* @param l [write] partial derivative of loss w.r.t. f
*/
void CNDCGRankLoss::LossAndGrad(Scalar& loss, TheMatrix& f, TheMatrix& l)
{
// chteo: here we make use of the subset information
loss = 0.0;
l.Zero();
Scalar* f_array = f.Data();
for(int q=0; q < _data->NumOfSubset(); q++)
{
//cout << "q = "<< q <<endl;
int offset = _data->subset[q].startIndex;
int subsetsize = _data->subset[q].size;
current_ideal_pi = sort_vectors[q];
vector<double> b = bs[q];
//compute_coefficients(offset, subsetsize, y_array, current_ideal_pi, a, b);
//cout << "before finding permutation\n";
/* find the best permutation */
find_permutation(subsetsize, offset, a, b, c, f_array, pi);
//cout << "after finding permutation\n";
//cout << "before finding delta\n";
/* compute the loss */
double value;
delta(subsetsize, a, b, pi, value);
//cout << "before finding delta\n";
loss += value;
for (int i=0;i<subsetsize;i++){
loss = loss + c[i]*(get(f_array, offset, pi[i]) - get(f_array, offset, i));
}
for (int i=0;i<subsetsize;i++){
//add(l, offset, i, c[pi[i]] - c[i]);
add(l, offset, i, - c[i]);
add(l, offset, pi[i], c[i]);
}
}
}
void CGenericLoss::ComputeLossAndGradient(double& loss, TheMatrix& grad)
{
loss = 0;
grad.Zero();
TheMatrix &w = _model->GetW();
double* dat = w.Data();
double* raw_g = grad.Data();
{
double* resy;
double* resybar;
map<int,int> ybar;
resy = new double [data->dim()];
resybar = new double [data->dim()];
minimize(data->nodeFeatures, &(data->nodeLabels), data->edgeFeatures, dat, dat + data->nNodeFeatures, ybar, data->nNodeFeatures, data->nEdgeFeatures, data->lossPositive, data->lossNegative, data->indexEdge, NULL, 1, data->firstOrderResponses);
Phi(data->nodeFeatures, &(data->nodeLabels), data->edgeFeatures, data->nNodeFeatures, data->nEdgeFeatures, resy, resy + data->nNodeFeatures, data->indexEdge);
Phi(data->nodeFeatures, &ybar, data->edgeFeatures, data->nNodeFeatures, data->nEdgeFeatures, resybar, resybar + data->nNodeFeatures, data->indexEdge);
loss += LabelLoss(data->nodeLabels, ybar, data->lossPositive, data->lossNegative, LOSS);
for (int j = 0; j < (int) data->dim(); j ++)
{
loss += dat[j]*(resybar[j]-resy[j]);
raw_g[j] += (1.0/data->N)*(resybar[j]-resy[j]);
}
delete [] resy;
delete [] resybar;
}
loss = loss/data->N;
}
/** Compute loss and gradient
*/
void CSMMMulticlassLoss::ComputeLossAndGradient(double& loss, TheMatrix& grad)
{
iterNum ++;
TheMatrix &w = _model->GetW();
loss = 0;
grad.Zero();
TheMatrix g(grad, SML::DENSE);
const vector<CSeqMulticlassLabel::seqlabel_struct> &Y = _data->labels();
const vector<CSeqMulticlassFeature::seqfeature_struct> &X = _data->features();
unsigned int trainExNum = 0;
vector <int > cvmark = _data->Getcvmark();
for(unsigned int i=0; i < m; i++)
{
if(cvmark.size()!=0)
{
if(cvmark[i]!=SMM::TRAIN_DATA)
continue;
}
trainExNum ++;
//if(cvmark)
vector<unsigned int> ybar(X[i].len,0);
vector<unsigned int> ybarlabel(X[i].len,0);
double labelloss = 0;
double marginloss = 0;
double w_dot_g = 0.0;;
// find best label y' and return the score wrt to y'
if(verbosity>=2)
{
cout <<"ex:"<< i<< endl;fflush(stdout);
}
if(is_single_action_persequence)
find_best_label_grammer(Y[i].pos,Y[i].type, X[i], w, ybar, ybarlabel, marginloss, labelloss, 0, _data->getNumOfClass());
else
find_best_label(Y[i].pos,Y[i].type, X[i], w, ybar, ybarlabel, marginloss, labelloss, 0, _data->getNumOfClass());
double labelloss_y = 0;
double marginloss_y = 0;
double labelloss_ybar = 0;
double marginloss_ybar = 0;
ComputeLoss(Y[i].pos,Y[i].type,ybar,ybarlabel,X[i],w,marginloss_ybar,labelloss_ybar,1);
if(lossw[0]!=0)
labelloss+=lossw[0];
if(lastDuration>0)
{
marginloss = marginloss_ybar;
labelloss = labelloss_ybar;
}
if(verbosity>=3)
{
ComputeLoss(Y[i].pos,Y[i].type,Y[i].pos,Y[i].type,X[i],w,marginloss_y,labelloss_y,1);
printf("dp------marginloss:%2.4f---labelloss:%2.4f------\n",marginloss,labelloss);
printf("ybar----marginloss:%2.4f---labelloss:%2.4f------\n",marginloss_ybar,labelloss_ybar);
printf("y-------marginloss:%2.4f---labelloss:%2.4f------\n",marginloss_y,labelloss_y);
if(abs(labelloss_ybar-labelloss)>1e-5)
{
printf("labelloss doesn't match!\n");
//exit(0);
}
if(abs(marginloss_ybar-marginloss)>1e-5)
{
printf("marginloss_ybar_dp:%2.4f != marginloss_ybar_computeLoss:%2.4f\n",marginloss,marginloss_ybar);
printf("marginloss doesn't match!\n");
}
}
// construct the gradient vector for the part of true y
const vector<unsigned int> &y = Y[i].pos;
const vector<unsigned int> &ylabel = Y[i].type;
g.Zero();
for(unsigned int j=0; j < y.size(); j++)
{
//g.Add(*(X[i].phi_1[y[j]]));
//g.Add(*(X[i].phi_2[y[j-1]][y[j]-y[j-1]-1]));
_data->TensorPhi1(X[i].phi_1[y[j]],ylabel[j],0,tphi_1);
g.Add(*tphi_1);
if(j > 0)
{
_data->TensorPhi2(X[i].phi_2[y[j-1]][y[j]-y[j-1]-1], ylabel[j-1], ylabel[j], 0,0,tphi_2);
g.Add(*tphi_2);
}
}
if(y.size() > 0)
{
//g.Add(*(X[i].phi_2[y[y.size()-1]][X[i].len-1 - y[y.size()-1]-1]));////
_data->TensorPhi2(X[i].phi_2[y[y.size()-1]][X[i].len - y[y.size()-1]-1 ], ylabel[y.size()-1], 0,0,0,tphi_2);
g.Add(*tphi_2);
}
// for predicted y'
for(unsigned int j=0; j < ybar.size(); j++)
{
//grad.Add(*(X[i].phi_1[ybar[j]]));
//grad.Add(*(X[i].phi_2[ybar[j-1]][ybar[j]-ybar[j-1]-1]));
_data->TensorPhi1(X[i].phi_1[ybar[j]],ybarlabel[j],0,tphi_1);
grad.Add(*tphi_1);
if(j>0)
{
_data->TensorPhi2(X[i].phi_2[ybar[j-1]][ybar[j]-ybar[j-1]-1], ybarlabel[j-1], ybarlabel[j], 0,0,tphi_2);
grad.Add(*tphi_2); ////
}
}
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[i].phi_2[ybar[ybar.size()-1]][X[i].len - ybar[ybar.size()-1]-1 ], ybarlabel[ybar.size()-1], 0, 0,0,tphi_2);
grad.Add(*tphi_2);
}
grad.Minus(g);
// accumulate the loss
w.Dot(g, w_dot_g);
loss = loss - w_dot_g + marginloss + labelloss;
}
scalingFactor = 1.0/trainExNum;
grad.Scale(scalingFactor);
loss *= scalingFactor;
if(verbosity)
{
double gnorm = 0.0;
grad.Norm2(gnorm);
cout << "gradient norm=" << gnorm << endl;
}
//Evaluate(_model);
}