void onHit(ObjectID id)
{
FeatureIndex* index = Registry::objectIndex()->get<FeatureIndex>( id );
Feature* feature = index ? index->getFeature( id ) : 0L;
if ( feature && feature->getFID() != _lastFID )
{
_grid->clearControls();
unsigned r=0;
_grid->setControl( 0, r, new LabelControl("FID", Color::Red) );
_grid->setControl( 1, r, new LabelControl(Stringify()<<feature->getFID(), Color::White) );
++r;
const AttributeTable& attrs = feature->getAttrs();
for( AttributeTable::const_iterator i = attrs.begin(); i != attrs.end(); ++i, ++r )
{
_grid->setControl( 0, r, new LabelControl(i->first, 14.0f, Color::Yellow) );
_grid->setControl( 1, r, new LabelControl(i->second.getString(), 14.0f, Color::White) );
}
if ( !_grid->visible() )
_grid->setVisible( true );
_lastFID = feature->getFID();
}
}
void onHit(ObjectID id)
{
// First see whether it's a feature:
FeatureIndex* index = Registry::objectIndex()->get<FeatureIndex>( id );
Feature* feature = index ? index->getFeature( id ) : 0L;
if ( feature )
{
s_fidLabel->setText( Stringify() << "Feature ID = " << feature->getFID() << " (oid = " << id << ")" );
s_nameLabel->setText( Stringify() << "Name = " << feature->getString("name") );
}
else
{
// Check whether it's an annotation:
AnnotationNode* anno = Registry::objectIndex()->get<AnnotationNode>( id );
if ( anno )
{
s_fidLabel->setText( Stringify() << "ObjectID = " << id );
s_nameLabel->setName( Stringify() << "Name = " << anno->getName() );
}
// None of the above.. clear.
else
{
s_fidLabel->setText( Stringify() << "oid = " << id );
s_nameLabel->setText( "Name = " );
}
}
s_highlightUniform->set( id );
}
void GBTree::train(const Dataset& trainset, int nDims, const Dataset& testset, const Dataset& valset, const MartParam& martParam, const Metric* m) {
this->nDims = nDims;
this->learningRate = martParam.learningRate;
doubles_t residues(trainset.size());
ints_t ind(trainset.size());
size_t nSamples = static_cast<size_t>(floor(trainset.size()*martParam.samplePrec+eps));
if(nSamples>trainset.size()) nSamples = trainset.size();
FVectors x(nSamples), tx(trainset.size());
for(size_t i=0;i<trainset.size();i++) tx[i] = trainset[i].x;
doubles_t weight(nSamples), y(nSamples), yhat(trainset.size()), ty(testset.size()), vy(valset.size()), fity(nSamples);
b = 0.0;
for(size_t i=0;i<trainset.size();i++) b+= trainset[i].y;
b/= trainset.size();
for(size_t i=0;i<trainset.size();i++) residues[i] = trainset[i].y-b;
for(size_t i=0;i<ind.size();i++) ind[i] = i;
double bestT = -numeric_limits<double>::max();
double bestV = -numeric_limits<double>::max();
double bestTV = -numeric_limits<double>::max();
Cart<FVector, doubles_t, doubles_t, doubles_t> cart(martParam.cartParam);
FeatureIndex *find = buildFeatureIndex(tx, nDims);
for(int iter=0;iter<martParam.nIters;iter++) {
random_shuffle(ind.begin(), ind.end());
for(size_t i=0;i<nSamples;i++) {
x[i] = trainset[ind[i]].x;
y[i] = residues[ind[i]];
weight[i] = 1.0;
// if(dcmp(y[i]-1)==0) weight[i] = 1.0; else weight[i] = 1e-2;
}
FeatureIndex *sampleInd = find->getSample(ints_t(ind.begin(), ind.begin()+nSamples));
#if 0
for(int p=0;p<nDims;p++) {
FeatureComp<FVector> fc(p, x), fc2(p, tx);
const int* pindAll = find->getDim(p);
assert(__gnu_cxx::is_sorted(pindAll, pindAll+tx.size(), fc2));
cout << "p = " << p << endl;
const int* pind = sampleInd->getDim(p);
cout << endl;
for(size_t i=1;i<nSamples;i++) {
if(fc(pind[i], pind[i-1])) {
for(size_t j=0;j<tx.size();j++) {
if(pindAll[j]==ind[pind[i-1]]) {
cout << " " << pindAll[j] << " " << pind[i-1] << " " << tx[pindAll[j]][p] << endl;
}
if(pindAll[j]==ind[pind[i]]) {
cout << " " << pindAll[j] << " " << pind[i] << " " << tx[pindAll[j]][p] << endl;
}
}
cout << "i = " << i << " " << pind[i-1] << " " << pind[i] << " " << x[pind[i-1]][p] << " " << x[pind[i]][p] << endl;
}
}
assert(__gnu_cxx::is_sorted(pind, pind+nSamples, fc));
}
#endif
treeVec.push_back(cart.fit(x, nDims, weight, y, fity, *sampleInd));
delete sampleInd;
double w1 = 0.0, w2 = 0.0;
for(size_t i=0;i<trainset.size();i++) {
yhat[i] = apply_cart(trainset[i].x, treeVec[iter]);
w1 += yhat[i]*residues[i]; w2 += yhat[i]*yhat[i];
}
assert(dcmp(w2)>0);
w.push_back(w1/w2);
// w.push_back(1.0);
// cout << w1/w2 << endl;
for(size_t i=0;i<trainset.size();i++) {
residues[i]-=yhat[i]*learningRate*w[iter];
yhat[i] = trainset[i].y-residues[i];
}
if(iter%10==0) {
test(trainset, yhat);
// doubles_t ndcg, ndcgT, ndcgV;
// cal_ndcg(trainset, yhat, ndcg);
double mtrain = m->evaluate(trainset, yhat);
cout << "Iter = " << iter << " Loss = " << getLoss(trainset, yhat) << " NDCG5 = " << mtrain;
if(!testset.empty()) {
test(testset, ty);
// cal_ndcg(testset, ty, ndcgT);
double mtest = m->evaluate(testset, ty);
cout << " TLoss = " << getLoss(testset, ty) << " TNDCG5 = " << mtest;
if(iter>100 && dcmp(bestT-mtest)<0) bestT = mtest;
if(!valset.empty()) {
test(valset, vy);
// cal_ndcg(valset, vy, ndcgV);
double mval = m->evaluate(valset, vy);
cout << " VLoss = " << getLoss(valset, vy) << " VNDCG5 = " << mval;
if(iter>100 && dcmp(bestV-mval)<0) {
bestV = mval; bestTV = mtest;
}
}
}
cout << endl;
}
}
delete find;
cout << "BestTest = " << bestT << endl;
cout << "BestVal = " << bestV << endl;
cout << "BestTestonVal = " << bestTV << endl;
}
