double CrossEntropy::operator() ( const dlib::matrix<double,0,1>& arg) const
{
double total = 0.0;
double n = 0.0;
std::vector<float> weight_vector (m_model->m_numModels);
for (int i=0; i < arg.nr(); i++) {
weight_vector[i] = arg(i);
}
if (m_model->m_mode == "interpolate") {
weight_vector = m_model->normalizeWeights(weight_vector);
}
for ( std::vector<multiModelStatisticsOptimization*>::const_iterator iter = m_optimizerStats.begin(); iter != m_optimizerStats.end(); ++iter ) {
multiModelStatisticsOptimization* statistics = *iter;
size_t f = statistics->f;
double score;
score = std::inner_product(statistics->p[m_iFeature].begin(), statistics->p[m_iFeature].end(), weight_vector.begin(), 0.0);
total -= (FloorScore(TransformScore(score))/TransformScore(2))*f;
n += f;
}
return total/n;
}
double CrossEntropyCounts::operator() ( const dlib::matrix<double,0,1>& arg) const
{
double total = 0.0;
double n = 0.0;
std::vector<float> weight_vector (m_model->m_numModels);
for (int i=0; i < arg.nr(); i++) {
weight_vector[i] = arg(i);
}
if (m_model->m_mode == "interpolate") {
weight_vector = m_model->normalizeWeights(weight_vector);
}
for ( std::vector<multiModelCountsStatisticsOptimization*>::const_iterator iter = m_optimizerStats.begin(); iter != m_optimizerStats.end(); ++iter ) {
multiModelCountsStatisticsOptimization* statistics = *iter;
size_t f = statistics->f;
double score;
if (m_iFeature == 0) {
score = m_model->m_combineFunction(statistics->fst, statistics->ft, weight_vector);
} else if (m_iFeature == 1) {
score = m_model->ComputeWeightedLexicalTranslationFromCache(statistics->lexCachee2f, weight_vector);
} else if (m_iFeature == 2) {
score = m_model->m_combineFunction(statistics->fst, statistics->fs, weight_vector);
} else if (m_iFeature == 3) {
score = m_model->ComputeWeightedLexicalTranslationFromCache(statistics->lexCachef2e, weight_vector);
} else {
score = 0;
UTIL_THROW(util::Exception, "Trying to optimize feature that I don't know. Aborting");
}
total -= (FloorScore(TransformScore(score))/TransformScore(2))*f;
n += f;
}
return total/n;
}
vector<float> PhraseDictionaryMultiModel::Optimize(OptimizationObjective *ObjectiveFunction, size_t numModels)
{
dlib::matrix<double,0,1> starting_point;
starting_point.set_size(numModels);
starting_point = 1.0;
try {
dlib::find_min_bobyqa(*ObjectiveFunction,
starting_point,
2*numModels+1, // number of interpolation points
dlib::uniform_matrix<double>(numModels,1, 1e-09), // lower bound constraint
dlib::uniform_matrix<double>(numModels,1, 1e100), // upper bound constraint
1.0, // initial trust region radius
1e-5, // stopping trust region radius
10000 // max number of objective function evaluations
);
} catch (dlib::bobyqa_failure& e) {
cerr << e.what() << endl;
}
vector<float> weight_vector (numModels);
for (int i=0; i < starting_point.nr(); i++) {
weight_vector[i] = starting_point(i);
}
cerr << "Cross-entropy: " << (*ObjectiveFunction)(starting_point) << endl;
return weight_vector;
}
