bool load(const char * model_file) {
if (!ltp::sdparser::NeuralNetworkParser::load(std::string(model_file))) {
return false;
}
setup_system();
build_feature_space();
return true;
}
void Postagger::train(void) {
const char * train_file = train_opt.train_file.c_str();
// read in training instance
read_instance(train_file);
TRACE_LOG("Read in [%d] instances.", train_dat.size());
model = new Model;
// build tag dictionary, map string tag to index
TRACE_LOG("Start build configuration");
build_configuration();
TRACE_LOG("Build configuration is done.");
TRACE_LOG("Number of labels: [%d]", model->labels.size());
// build feature space from the training instance
TRACE_LOG("Start building feature space.");
build_feature_space();
TRACE_LOG("Building feature space is done.");
TRACE_LOG("Number of features: [%d]", model->space.num_features());
model->param.realloc(model->space.dim());
TRACE_LOG("Allocate [%d] dimensition parameter.", model->space.dim());
PostaggerWriter writer(cout);
if (train_opt.algorithm == "mira") {
// use mira algorithm
/*kbest_decoder = new KBestDecoder(L);
for (int iter = 0; iter < train_opt.max_iter; ++ iter) {
for (int i = 0; i < train_dat.size(); ++ i) {
extract_features(train_dat[i]);
calculate_scores(train_dat[i]);
KBestDecoder::KBestDecodeResult result;
kbest_decoder->decode(train_dat[i], result);
}
}*/
} else {
// use pa or average perceptron algorithm
decoder = new Decoder(model->num_labels());
TRACE_LOG("Allocated plain decoder");
for (int iter = 0; iter < train_opt.max_iter; ++ iter) {
TRACE_LOG("Training iteraition [%d]", (iter + 1));
for (int i = 0; i < train_dat.size(); ++ i) {
// extract_features(train_dat[i]);
Instance * inst = train_dat[i];
calculate_scores(inst, false);
decoder->decode(inst);
if (inst->features.dim() == 0) {
collect_features(inst, inst->tagsidx, inst->features);
}
collect_features(inst, inst->predicted_tagsidx, inst->predicted_features);
// writer.debug(inst, true);
if (train_opt.algorithm == "pa") {
SparseVec update_features;
update_features.zero();
update_features.add(train_dat[i]->features, 1.);
update_features.add(train_dat[i]->predicted_features, -1.);
double error = train_dat[i]->num_errors();
double score = model->param.dot(update_features, false);
double norm = update_features.L2();
double step = 0.;
if (norm < EPS) {
step = 0;
} else {
step = (error - score) / norm;
}
model->param.add(update_features,
iter * train_dat.size() + i + 1,
step);
} else if (train_opt.algorithm == "ap") {
SparseVec update_features;
update_features.zero();
update_features.add(train_dat[i]->features, 1.);
update_features.add(train_dat[i]->predicted_features, -1.);
model->param.add(update_features,
iter * train_dat.size() + i + 1,
1.);
}
if ((i+1) % train_opt.display_interval == 0) {
TRACE_LOG("[%d] instances is trained.", i+1);
}
}
TRACE_LOG("[%d] instances is trained.", train_dat.size());
model->param.flush( train_dat.size() * (iter + 1) );
Model * new_model = truncate();
swap(model, new_model);
evaluate();
std::string saved_model_file = (train_opt.model_name + "." + strutils::to_str(iter) + ".model");
std::ofstream ofs(saved_model_file.c_str(), std::ofstream::binary);
swap(model, new_model);
new_model->save(ofs);
delete new_model;
// model->save(ofs);
TRACE_LOG("Model for iteration [%d] is saved to [%s]",
iter + 1,
saved_model_file.c_str());
}
}
}
