void Config::get_layers_config(string &str,
SoftMax &SMR) {
std::vector<string> layers;
if (str.empty())
return;
int head = 0;
int tail = 0;
while (1) {
if (head == str.length())
break;
if (str[head] == '$') {
tail = head + 1;
while (1) {
if (tail == str.length())
break;
if (str[tail] == '&')
break;
++tail;
}
string sub = str.substr(head, tail - head + 1);
if (sub[sub.length() - 1] == '&') {
sub.erase(sub.begin() + sub.length() - 1);
sub.erase(sub.begin());
layers.push_back(sub);
}
str.erase(head, tail - head + 1);
} else
++head;
}
for (int i = 0; i < layers.size(); i++) {
int type = get_word_type(layers[i], "LAYER");
switch (type) {
case 0: {
int nn = get_word_int(layers[i], "NUM_HIDDEN_NEURONS");
double wd = get_word_float(layers[i], "WEIGHT_DECAY");
double dr = get_word_float(layers[i], "DROPOUT_RATE");
HiddenConfigs.push_back(HiddenConfig(nn, wd, dr));
break;
}
case 1: {
int classnum = get_word_int(layers[i], "NUM_CLASSES");
double wd = get_word_float(layers[i], "WEIGHT_DECAY");
SMR.set_NumClasses(classnum);
SMR.set_WeightDecay(wd);
break;
}
}
}
}
void Config::init(string path,
SoftMax &SMR) {
m_configStr = read_2_string(path);
deleteComment();
deleteSpace();
get_layers_config(m_configStr,SMR);
use_log = get_word_bool(m_configStr, "USE_LOG");
batch_size = get_word_int(m_configStr, "BATCH_SIZE");
non_linearity = get_word_type(m_configStr, "NON_LINEARITY");
training_epochs = get_word_int(m_configStr, "TRAINING_EPOCHS");
lrate_w = get_word_float(m_configStr, "LRATE_W");
lrate_b = get_word_float(m_configStr, "LRATE_B");
iter_per_epo = get_word_int(m_configStr, "ITER_PER_EPO");
ngram = get_word_int(m_configStr, "NGRAM");
training_percent = get_word_float(m_configStr, "TRAINING_PERCENT");
cout
<< "****************************************************************************"
<< endl
<< "** READ CONFIG FILE COMPLETE "
<< endl
<< "****************************************************************************"
<< endl << endl;
for (int i = 0; i < HiddenConfigs.size(); i++) {
cout << "***** hidden layer: " << i << " *****" << endl;
cout << "NumHiddenNeurons = " << HiddenConfigs[i].get_NeuronNum()
<< endl;
cout << "WeightDecay = " << HiddenConfigs[i].get_WeightDecay() << endl;
cout << "DropoutRate = " << HiddenConfigs[i].get_DropoutRate() << endl << endl;
}
cout << "***** softmax layer: *****" << endl;
// cout<<"NumClasses = "<<softmaxConfig.NumClasses<<endl;
cout << "WeightDecay = " << SMR.get_WeightDecay() << endl << endl;
cout << "***** general config *****" << endl;
cout << "use_log = " << use_log << endl;
cout << "batch size = " << batch_size << endl;
cout << "non-linearity method = " << non_linearity << endl;
cout << "training epochs = " << training_epochs << endl;
cout << "learning rate for weight matrices = " << lrate_w << endl;
cout << "learning rate for bias = " << lrate_b << endl;
cout << "iteration per epoch = " << iter_per_epo << endl;
cout << "ngram = " << ngram << endl;
cout << "training percent = " << training_percent << endl;
cout << endl;
}
void FeatureGenBackoff::backed_off_features(const Sentence &sentence,
const AppliedRule &rule,
int index,
int extra,
FeatureState *state) const {
int word = get_word_int(lexicon_, sentence, index);
int rule_ = rule.rule;
int X = grammar_->head_symbol[rule.rule];
state->inc(word, rule_, extra);
state->inc(word, X, extra);
if (!simple_) {
int tag = get_tag_int(lexicon_, sentence, index);
state->inc(tag, rule_, extra);
state->inc(tag, X, extra);
}
}
