Examples<EXAMPLE>::Examples(const Examples<EXAMPLE2>& e) {
Debug::log(2) << "Examples<" << EXAMPLE::name() << ">::Examples(Examples<" << EXAMPLE2::name() << "> [" << e.size() << "])...\n";
Debug::log(3) << stats::resource_usage() << "\n";
this->reserve(e.size());
this->insert(this->begin(), e.begin(), e.end());
assert(this->size() == e.size());
Debug::log(2) << "...Examples<" << EXAMPLE::name() << ">::Examples(Examples<" << EXAMPLE2::name() << "> [" << e.size() << "])\n";
Debug::log(2) << stats::resource_usage() << "\n";
}
/// Sanity check that the initial Weights computed from exmpls
/// match those read from the hypothesis file in this Tree.
/// \todo It would be much more efficient if we move this to
/// Ensemble::set_confidences.
template<typename EXAMPLE> void Tree::verify_initial_weights(const Examples<EXAMPLE>& exmpls) const {
hash_map<NodeID, Weights> initial_weights;
for(typename Examples<EXAMPLE>::const_iterator e = exmpls.begin(); e != exmpls.end(); e++) {
const Node& n = *this->find_leaf(*e);
initial_weights[n.id()].add(e->initial_weight(), e->is_correct());
}
for (vector<Node>::const_iterator n = nodes.begin(); n != nodes.end(); n++) {
if (!n->is_leaf()) continue;
// assert(fabs(n->initial_weights().pos()() - initial_weights[n->id()].pos()()) < parameter::small_epsilon());
// assert(fabs(n->initial_weights().neg()() - initial_weights[n->id()].neg()()) < parameter::small_epsilon());
if (fabs(n->initial_weights().pos()() - initial_weights[n->id()].pos()()) > parameter::small_epsilon()) {
ostringstream o;
o << "|read initial posweight - calc initial posweight| = " << fabs(n->initial_weights().pos()() - initial_weights[n->id()].pos()());
Debug::warning(__FILE__, __LINE__, o.str());
}
if (fabs(n->initial_weights().neg()() - initial_weights[n->id()].neg()()) > parameter::small_epsilon()) {
ostringstream o;
o << "|read initial negweight - calc initial negweight| = " << fabs(n->initial_weights().neg()() - initial_weights[n->id()].neg()());
Debug::warning(__FILE__, __LINE__, o.str());
}
}
}
/// Weight the leaves of this tree, and update the Example weights.
/// \todo Backprune splits that don't reduce loss, and backprune leaves
/// that don't have enough weight/exmpls to meet the initial splitting
/// criteria
/// \todo Weight the internal nodes too, for debugging purposes?
template<typename EXAMPLE> void Tree::weight_leaves_and_update_examples(Examples<EXAMPLE>& exmpls) {
Debug::log(1) << "\nTree::weight_leaves_and_update_examples(Examples<" << EXAMPLE::name() << ">)...\n";
vector<Node>::iterator n;
Double orig_total_weight;
hash_map<NodeID, set<ID<Sentence> > > sentences;
hash_map<NodeID, ExamplePtrs<EXAMPLE> > leaves;
unsigned totcnt = 0;
for(typename Examples<EXAMPLE>::iterator ex = exmpls.begin(); ex != exmpls.end(); ex++) {
// Find the node that e falls into.
const Node* n = this->find_leaf(*ex);
assert(n->is_leaf());
assert(n->id() != NO_NODE);
sentences[n->id()].insert(ex->sentence());
sentences[NO_NODE].insert(ex->sentence());
leaves[n->id()].push_back(&(*ex));
// FIXME: This won't work if there's noise
orig_total_weight += ex->weight();
totcnt++;
if (totcnt % 100000 == 0)
Debug::log(3) << "\tProcessed " << totcnt << " examples in Tree::weight_leaves()\n";
if (totcnt % 10000 == 0)
Debug::log(4) << "\tProcessed " << totcnt << " examples in Tree::weight_leaves()\n";
}
Debug::log(2) << "Done processing " << totcnt << " examples in Tree::weight_leaves()\n";
// Compute the confidence for each leaf.
unsigned leafcnt = 0;
unsigned sentence_cnt = 0;
unsigned example_cnt = 0;
for (n = nodes.begin(); n != nodes.end(); n++) {
if (n->is_leaf()) {
assert(leaves.find(n->id()) != leaves.end());
const ExamplePtrs<EXAMPLE>& leaf_examples = leaves.find(n->id())->second;
double orig_unpenalized_loss = leaf_examples.unpenalized_loss();
Weights initial_weight = leaf_examples.initial_weight();
example_cnt += leaf_examples.size();
n->set_confidence_and_initial_weights(leaf_examples);
// Add the leaf confidence to the leaf Examples.
leaves.find(n->id())->second.add_confidence(n->confidence());
// Update the confidence of this leaf's Example%s.
double unpenalized_loss = leaf_examples.unpenalized_loss();
leafcnt++;
// Debug::log(2) << "Weighted leaf:\n";
// Debug::log(2) << n->to_string("\t");
Debug::log(2) << n->to_string();
assert(sentences.find(n->id()) != sentences.end());
sentence_cnt += sentences.find(n->id())->second.size();
Debug::log(2) << "\t" << leaf_examples.size() << " examples from ";
Debug::log(2) << sentences.find(n->id())->second.size() << " different sentences\n";
double penalty = n->penalty();
Debug::log(2) << "\tloss = " << unpenalized_loss + penalty << " = " << unpenalized_loss << " (unpenalized loss) + " << penalty << " (penalty)";
if (n->confidence() != 0) Debug::log(2) << " (conf=0 loss was " << orig_unpenalized_loss << ")";
Debug::log(2) << "\n";
}
}
assert(example_cnt == exmpls.size());
Debug::log(2) << "Examples from " << sentences[NO_NODE].size() << " different sentences.\n";
Debug::log(2) << "Done weighting " << leafcnt << " leaves in Tree::weight_leaves()\n";
Debug::log(2) << stats::resource_usage() << "\n";
// WRITEME: Backprune splits that don't reduce loss
Debug::log(1) << "...Tree::weight_leaves_and_update_examples(Examples<" << EXAMPLE::name() << ">)\n";
}