void ISimpleSubtractorImpl::Train(const unsigned char *in_img,
const unsigned int in_num_rows, const unsigned int in_num_cols,
const unsigned int in_img_step)
{
m_num_rows = in_num_rows;
m_num_cols = in_num_cols;
AllocateModel();
InitializeModel(in_img, in_img_step);
}
void FNeuralNetLMBase::TrainLM(const string &validationfile,
const string &outbase,
bool nce_ppl) {
// =============
// Prepare for the training
// Equivalent to ReadLM
word_vocab_.ReadVocabFromTxt(word_vocab_filename_);
if (word_vocab_.empty()) {
cerr << "empty word vocabulary!" << endl;
exit(EXIT_FAILURE);
}
factor_vocab_.ReadVocabFromTxt(factor_vocab_filename_);
if (factor_vocab_.empty()) {
cerr << "empty factor vocabulary!" << endl;
exit(EXIT_FAILURE);
}
ReadDecompFromTxt(decomp_filename_);
PrintParams();
CheckParams();
AllocateModel();
InitializeNeuralNet();
// ==== END ====
// Read the data
FNNLMDataReader train_data(train_filenames_, &word_vocab_, &factor_vocab_,
shuffle_datafiles_, shuffle_sentences_);
vector<string> validation_filenames = { validationfile };
FNNLMDataReader validation_data(validation_filenames, &word_vocab_, &factor_vocab_, false, false);
// Set NCE sampling.
if (nce_) {
// TODO: flatten noise_distribution_?
vector<int> word_count(word_vocab_.size(), 0);
int num_word_tokens = 0;
const size_t eos_widx = word_vocab().eos_idx();
vector<int> factor_count(factor_vocab_.size(), 0);
int num_factor_tokens = 0;
const size_t eos_fidx = factor_vocab().eos_idx();
vector<pair<size_t, vector<size_t>>> sentence;
train_data.StartEpoch();
while(train_data.GetSentence(sentence)) {
for (vector<pair<size_t, vector<size_t>>>::const_iterator it = sentence.begin(); it != sentence.end(); ++it) {
word_count[it->first]++;
num_word_tokens++;
if (weight_factor_output_ > 0) {
for (size_t p = 0; p < it->second.size(); p++) {
factor_count[it->second[p]]++;
num_factor_tokens++;
}
}
}
word_count[eos_widx]++;
num_word_tokens++;
if (weight_factor_output_ > 0) {
factor_count[eos_fidx]++;
num_factor_tokens++;
}
}
word_noise_distribution_ = Distribution(word_count.begin(), word_count.end());
word_noise_pdf_ = word_noise_distribution_.param().probabilities();
if (weight_factor_output_ > 0) {
factor_noise_distribution_ = Distribution(factor_count.begin(), factor_count.end());
factor_noise_pdf_ = factor_noise_distribution_.param().probabilities();
}
NCECheckSampling();
log_num_negative_samples_ = log(num_negative_samples_);
}
BatchSGDTrain(train_data, validation_data, outbase, nce_ppl);
cout << "================================================================================" << endl;
cout << "Log-likelihood (base e) on validation is: " \
<< EvalLM(validation_data, false) << endl;
}
