void
PerceptronModel::DefaultUpdater::Update(Model *m, CandidateSet &example) {
PerceptronModel *model = dynamic_cast<PerceptronModel *>(m);
++(model->num_updates_);
unordered_set<int> gold_features;
unordered_set<int> best_scoring_features;
model->ComputeFeaturesToUpdate(example, gold_features, best_scoring_features);
model->models_.UpdateGoldAndCandidateFeatureAverages(model->time_,
gold_features,
best_scoring_features);
double step_size =
model->ComputeStepSize(gold_features, best_scoring_features, example);
// Finally, update perceptrons.
if (DEBUG >= 2) {
cerr << "Updating weights for gold features [";
for (unordered_set<int>::const_iterator it = gold_features.begin();
it != gold_features.end(); ++it) {
cerr << " " << *it;
}
cerr << "] from\n\t" << example.GetGold() << endl;
cerr << "Updating weights for best scoring features [";
for (unordered_set<int>::const_iterator it = best_scoring_features.begin();
it != best_scoring_features.end(); ++it) {
cerr << " " << *it;
}
cerr << "] from\n\t" << example.GetBestScoring() << endl;
}
double positive_step = step_size;
model->models_.UpdateWeights(model->time_, gold_features,
example.GetGold().features(), positive_step);
double negative_step = -step_size;
model->models_.UpdateWeights(model->time_, best_scoring_features,
example.GetBestScoring().features(), negative_step);
if (DEBUG >=2) {
cerr << "Raw model: " << model->models_.GetModel(true) << endl;
cerr << "Avg model: " << model->models_.GetModel(false) << endl;
}
}
void
PerceptronModel::ComputeFeaturesToUpdate(const CandidateSet &example,
unordered_set<int> &
gold_features_to_update,
unordered_set<int> &
best_scoring_features_to_update)
const {
// Collect gold features that are not in best-scoring candidate.
const FeatureVector<int,double> &gold_features =
example.GetGold().features();
const FeatureVector<int,double> &best_scoring_features =
example.GetBestScoring().features();
if (DEBUG >= 2) {
cerr << "Gold index: " << example.gold_index()
<< "; best scoring index: " << example.best_scoring_index()
<< endl;
cerr << "Original gold features: " << gold_features << endl
<< "Original best scoring features: " << best_scoring_features << endl;
}
gold_features.GetNonZeroFeatures(gold_features_to_update);
best_scoring_features.RemoveEqualFeatures(gold_features,
gold_features_to_update);
if (DEBUG >= 2) {
cerr << "Time:" << time_.to_string() << ": new gold features: [";
for (unordered_set<int>::const_iterator it =
gold_features_to_update.begin();
it != gold_features_to_update.end();
++it) {
cerr << " " << *it;
}
cerr << "]" << endl;
}
best_scoring_features.GetNonZeroFeatures(best_scoring_features_to_update);
gold_features.RemoveEqualFeatures(best_scoring_features,
best_scoring_features_to_update);
if (DEBUG >= 2) {
cerr << "Time:" << time_.to_string() << ": new best scoring features: [";
for (unordered_set<int>::const_iterator it =
best_scoring_features_to_update.begin();
it != best_scoring_features_to_update.end();
++it) {
cerr << " " << *it;
}
cerr << "]" << endl;
}
}
