inline void BoostOneIter( const DMatrix &train,
float *grad, float *hess, size_t len, int bst_group ){
this->grad_.resize( len ); this->hess_.resize( len );
memcpy( &this->grad_[0], grad, sizeof(float)*len );
memcpy( &this->hess_[0], hess, sizeof(float)*len );
if( grad_.size() == train.Size() ){
if( bst_group < 0 ) bst_group = 0;
base_gbm.DoBoost(grad_, hess_, train.data, train.info.root_index, bst_group);
}else{
utils::Assert( bst_group == -1, "must set bst_group to -1 to support all group boosting" );
int ngroup = base_gbm.NumBoosterGroup();
utils::Assert( grad_.size() == train.Size() * (size_t)ngroup, "BUG: UpdateOneIter: mclass" );
std::vector<float> tgrad( train.Size() ), thess( train.Size() );
for( int g = 0; g < ngroup; ++ g ){
memcpy( &tgrad[0], &grad_[g*tgrad.size()], sizeof(float)*tgrad.size() );
memcpy( &thess[0], &hess_[g*tgrad.size()], sizeof(float)*tgrad.size() );
base_gbm.DoBoost(tgrad, thess, train.data, train.info.root_index, g );
}
}
}
int trustregion(NLP1* nlp, std::ostream *fout,
SymmetricMatrix& H, ColumnVector& search_dir,
ColumnVector& sx, real& TR_size, real& step_length,
real stpmax, real stpmin)
{
/****************************************************************************
* subroutine trustregion
*
* Purpose
* find a step which satisfies the Goldstein-Armijo line search conditions
*
* Compute the dogleg step
* Compute the predicted reduction, pred, of the quadratic model
* Compute the actual reduction, ared
* IF ared/pred > eta
* THEN x_vec = x_vec + d_vec
* TR_size >= TR_size
* Compute the gradient g_vec at the new point
* ELSE TR_size < ||d_vec||
*
* Parameters
* nlp --> pointer to nonlinear problem object
*
* search_dir --> Vector of length n which specifies the
* newton direction on input. On output it will
* contain the step
*
* step_length <-- is a nonnegative variable.
* On output step_length contains the step size taken
*
* ftol --> default Value = 1.e-4
* ftol should be smaller than 5.e-1
* suggested value = 1.e-4 for newton methods
* = 1.e-1 for more exact line searches
* xtol --> default Value = 2.2e-16
* gtol --> default Value = 0.9
* gtol should be greater than 1.e-4
*
* termination occurs when the sufficient decrease
* condition and the directional derivative condition are
* satisfied.
*
* stpmin and TR_size are nonnegative input variables which
* specify lower and upper bounds for the step.
* stpmin Default Value = 1.e-9
*
* Initial version Juan Meza November 1994
*
*
*****************************************************************************/
// Local variables
int n = nlp->getDim();
bool debug = nlp->getDebug();
bool modeOverride = nlp->getModeOverride();
ColumnVector tgrad(n), newton_dir(n), tvec(n), xc(n), xtrial(n);
real fvalue, fplus, dnorm;
real eta1 = .001;
real eta2 = .1;
real eta3 = .75;
real rho_k;
int iter = 0;
int iter_max = 100;
real dd1, dd2;
real ared, pred;
int dog_step;
real TR_MAX = stpmax;
static char *steps[] = {"C", "D", "N", "B"};
static bool accept;
//
// Initialize variables
//
fvalue = nlp->getF();
xc = nlp->getXc();
step_length = 1.0;
tgrad = nlp->getGrad();
newton_dir = search_dir;
if (debug) {
*fout << "\n***************************************";
*fout << "***************************************\n";
*fout << "\nComputeStep using trustregion\n";
*fout << "\tStep ||step|| ared pred TR_size \n";
}
while (iter < iter_max) {
iter++;
//
// Compute the dogleg step
//
search_dir = newton_dir;
dog_step = dogleg(nlp, fout, H, tgrad, search_dir, sx, dnorm, TR_size, stpmax);
step_length = dnorm;
//
// Compute pred = -g'd - 1/2 d'Hd
//
dd1 = Dot(tgrad,search_dir);
tvec = H * search_dir;
dd2 = Dot(search_dir,tvec);
pred = -dd1 - dd2/2.0;
//
// Compute objective function at trial point
//
xtrial = xc + search_dir;
if (modeOverride) {
nlp->setX(xtrial);
nlp->eval();
fplus = nlp->getF();
}
else
fplus = nlp->evalF(xtrial);
ared = fvalue - fplus;
//
// Should we take this step ?
//
rho_k = ared/pred;
accept = false;
if (rho_k >= eta1) accept = true;
//
// Update the trust region
//
if (accept) {
//
// Do the standard updating
//
if (rho_k <= eta2) {
// Model is just sort of bad
// New trust region will be TR_size/2
if (debug) {
*fout << "trustregion: rho_k = " << e(rho_k,14,4)
<< " eta2 = " << e(eta2,14,4) << "\n";
}
TR_size = step_length / 2.0;
if (TR_size < stpmin) {
*fout << "***** Trust region too small to continue.\n";
nlp->setX(xc);
nlp->setF(fvalue);
nlp->setGrad(tgrad);
return(-1);
}
}
else if ((eta3 <= rho_k) && (rho_k <= (2.0 - eta3))) {
// Model is PRETTY good
// Double trust region
if (debug) {
*fout << "trustregion: rho_k = " << e(rho_k,14,4)
<< " eta3 = " << e(eta3,14,4) << "\n";
}
TR_size = min(2.0*TR_size, TR_MAX);
}
else {
//
// All other cases
//
TR_size = max(2.0*step_length,TR_size);
TR_size = min(TR_size, TR_MAX);
if (debug) {
*fout << "trustregion: rho_k = " << e(rho_k,14,4) << "\n";
}
}
}
else {
// Model is REALLY bad
//
TR_size = step_length/10.0;
if (debug) {
*fout << "trustregion: rho_k = " << e(rho_k,14,4) << "n"
<< " eta1 = " << e(eta1,14,4) << "\n";
}
if (TR_size < stpmin) {
*fout << "***** Trust region too small to continue.\n";
nlp->setX(xc);
nlp->setF(fvalue);
nlp->setGrad(tgrad);
return(-1);
}
}
//
// Accept/Reject Step
//
if (accept) {
//
// Update x, f, and grad
//
if (!modeOverride) {
nlp->setX(xtrial);
nlp->setF(fplus);
nlp->evalG();
}
if (debug) {
*fout << "\t Step ||step|| ared pred TR_size \n";
*fout << "Accept " << steps[dog_step] << e(step_length,14,4)
<< e(ared,14,4) << e(pred,14,4) << e(TR_size,14,4) << "\n";
}
return(dog_step);
}
else {
//
// Reject step
//
if (debug) {
*fout << "\t Step ||step|| ared pred TR_size \n";
*fout << "Reject " << steps[dog_step] << e(step_length,14,4)
<< e(ared,14,4) << e(pred,14,4) << e(TR_size,14,4) << "\n";
}
}
}
nlp->setX(xc);
nlp->setF(fvalue);
nlp->setGrad(tgrad);
return(-1); // Too many iterations
}
