void CBMRM::DisplayAfterTrainingInfo(unsigned int iter, double finalExactObjVal,
double approxObjVal, double loss,
TheMatrix& w_best, CTimer& lossAndGradientTime,
CTimer& innerSolverTime, CTimer& totalTime)
{
// legends
if(verbosity >= 1)
{
printf("\n[Legends]\n");
if(verbosity > 1)
printf("pobj: primal objective function value"
"\naobj: approximate objective function value\n");
printf("gam: gamma (approximation error) "
"\neps: lower bound on gam "
"\nloss: loss function value "
"\nreg: regularizer value\n");
}
double norm1 = 0, norm2 = 0, norminf = 0;
w_best.Norm1(norm1);
w_best.Norm2(norm2);
w_best.NormInf(norminf);
printf("\nNote: the final w is the w_t where J(w_t) is the smallest.\n");
printf("No. of iterations: %d\n",iter);
printf("Primal obj. val.: %.6e\n",finalExactObjVal);
printf("Approx obj. val.: %.6e\n",approxObjVal);
printf("Primal - Approx.: %.6e\n",finalExactObjVal-approxObjVal);
printf("Loss: %.6e\n",loss);
printf("|w|_1: %.6e\n",norm1);
printf("|w|_2: %.6e\n",norm2);
printf("|w|_oo: %.6e\n",norminf);
// display timing profile
printf("\nCPU seconds in:\n");
printf("1. loss and gradient: %8.2f\n", lossAndGradientTime.CPUTotal());
printf("2. solver: %8.2f\n", innerSolverTime.CPUTotal());
printf(" Total: %8.2f\n", totalTime.CPUTotal());
printf("Wall-clock total: %8.2f\n", totalTime.WallclockTotal());
}
/** 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);
}