/**
* 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());
}
/**
* Compute NDCGRank 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.
*/
void CNDCGRankLoss::Loss(Scalar& loss, TheMatrix& f)
{
// chteo: here we make use of the subset information
loss = 0.0;
Scalar* f_array = f.Data();
for(int q=0; q < _data->NumOfSubset(); q++)
{
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);
/* find the best permutation */
find_permutation(subsetsize, offset, a, b, c, f_array, pi);
/* compute the loss */
double value;
delta(subsetsize, a, b, pi, value);
loss += value;
for (int i=0;i<subsetsize;i++){
loss = loss + c[i]*(get(f_array, offset, pi[i]) - get(f_array, offset, i));
}
//free(c);
//free(a);
//free(b);
//free(pi);
}
}
/**
* 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 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.
*/
void CLogisticLoss::Loss(double& loss, TheMatrix& f)
{
loss = 0;
f.ElementWiseMult(_data->labels()); // f = y*f
double* f_array = f.Data(); // pointer to memory location of f (faster element access)
int len = f.Length();
for(int i=0; i < len; i++)
{
if(fabs(f_array[i]) == 0.0)
loss += LN2;
else if (f_array[i] > 0.0)
loss += log(1+exp(-f_array[i]));
else
loss += log(1+exp(f_array[i])) - f_array[i];
}
}
/**
* 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;
}