void FidoInterface::gridSearch(std::vector<double>& alpha_search,
std::vector<double>& beta_search,
std::vector<double>& gamma_search) {
double gamma_best = -1.0, alpha_best = -1.0, beta_best = -1.0;
double best_objective = -100000000;
double current_objective;
for (unsigned int i = 0; i < gamma_search.size(); i++) {
double gamma_local = gamma_search[i];
for (unsigned int j = 0; j < alpha_search.size(); j++) {
double alpha_local = alpha_search[j];
for (unsigned int k = 0; k < beta_search.size(); k++) {
double beta_local = beta_search[k];
current_objective = calcObjective(alpha_local, beta_local, gamma_local);
if (current_objective > best_objective) {
best_objective = current_objective;
gamma_best = gamma_local;
alpha_best = alpha_local;
beta_best = beta_local;
}
}
}
}
alpha_ = alpha_best;
beta_ = beta_best;
gamma_ = gamma_best;
}
void dtw_model::validate(Corpus& corpus) {
static const size_t MINIATURE_SIZE = 10000;
static vector<tsample> samples = corpus.getSamples(MINIATURE_SIZE);
static double objective = 0;
static bool aboutToStop = false;
static const double SOFT_THRESHOLD = 2e-6 * _learning_rate;
static const double HARD_THRESHOLD = SOFT_THRESHOLD * 0.1;
static size_t MIN_ITERATION = 128;
static size_t itr = 0;
double obj = calcObjective(samples);
double diff = obj - objective;
double improveRate = abs(diff / objective);
printf("objective = %.7f \t prev-objective = %.7f \n", obj, objective);
printf("improvement rate on dev-set of size %lu = %.6e ", samples.size(), improveRate);
printf(", still "GREEN"%.0f"COLOREND" times of threshold \n", improveRate / SOFT_THRESHOLD);
if (itr > MIN_ITERATION) {
if (improveRate != improveRate)
exit(-1);
if (improveRate < HARD_THRESHOLD) {
printf("\nObjective function on dev-set is no longer decreasing...\n");
printf("Training process "GREEN"DONE"COLOREND"\n");
// doPause();
exit(0);
}
else if (aboutToStop || improveRate < SOFT_THRESHOLD) {
aboutToStop = true;
_learning_rate /= 2;
}
}
objective = obj;
++itr;
}
