void EvaluateModel(
const Config& config, const model& nn, propagator& prop_validation,
const shared_ptr<Corpus>& test_corpus, const shared_ptr<Vocabulary>& vocab,
double& current_learning_rate, double& current_validation_ll) {
if (test_corpus->size() > 0) {
double log_likelihood = 0.0;
Matrix<double,Dynamic,Dynamic> scores(vocab->size(), config.minibatch_size);
Matrix<double,Dynamic,Dynamic> output_probs(vocab->size(), config.minibatch_size);
MinibatchExtractor extractor(test_corpus, vocab, config);
cerr << endl;
cerr << "Validation minibatches: " << endl;
int num_batches = (test_corpus->size() - 1) / config.minibatch_size + 1;
for (int batch = 0; batch < num_batches; batch++) {
if (batch % 50 == 0) {
cerr << batch << "... ";
}
data_size_t start_index = config.minibatch_size * batch;
MatrixInt minibatch = extractor.extract(start_index);
prop_validation.fProp(minibatch.topRows(config.ngram_size - 1));
// Do full forward prop through output word embedding layer
start_timer(4);
prop_validation.output_layer_node.param->fProp(prop_validation.second_hidden_activation_node.fProp_matrix, scores);
stop_timer(4);
// And softmax and loss. Be careful of short minibatch
double minibatch_log_likelihood;
start_timer(5);
SoftmaxLogLoss().fProp(
scores.leftCols(minibatch.cols()),
minibatch.row(config.ngram_size - 1),
output_probs,
minibatch_log_likelihood);
stop_timer(5);
log_likelihood += minibatch_log_likelihood;
}
cerr << endl;
cerr << "Validation log-likelihood: " << log_likelihood << endl;
cerr << " perplexity: " << exp(-log_likelihood / test_corpus->size()) << endl;
// If the validation perplexity decreases, halve the learning rate.
if (current_validation_ll != 0 && log_likelihood < current_validation_ll) {
current_learning_rate /= 2;
} else {
current_validation_ll = log_likelihood;
if (config.model_output_file != "") {
cerr << "Writing model to " << config.model_output_file << endl;
ofstream fout(config.model_output_file);
nn.write(fout);
vocab->write(fout);
cerr << "Done writing model" << endl;
}
}
}
}
void compute_validation_perplexity(int ngram_size, int output_vocab_size, int validation_minibatch_size, int validation_data_size, int num_validation_batches, param & myParam, propagator & prop_validation, Map< Matrix<int,Dynamic,Dynamic> > & validation_data, double & current_learning_rate, double & current_validation_ll)
{
double log_likelihood = 0.0;
Matrix<double,Dynamic,Dynamic> scores(output_vocab_size, validation_minibatch_size);
Matrix<double,Dynamic,Dynamic> output_probs(output_vocab_size, validation_minibatch_size);
Matrix<int,Dynamic,Dynamic> minibatch(ngram_size, validation_minibatch_size);
for (int validation_batch =0;validation_batch < num_validation_batches;validation_batch++)
{
int validation_minibatch_start_index = validation_minibatch_size * validation_batch;
int current_minibatch_size = min(validation_minibatch_size,
validation_data_size - validation_minibatch_start_index);
minibatch.leftCols(current_minibatch_size) = validation_data.middleCols(validation_minibatch_start_index,
current_minibatch_size);
prop_validation.fProp(minibatch.topRows(ngram_size-1));
// Do full forward prop through output word embedding layer
start_timer(4);
if (prop_validation.skip_hidden)
prop_validation.output_layer_node.param->fProp(prop_validation.first_hidden_activation_node.fProp_matrix, scores);
else
prop_validation.output_layer_node.param->fProp(prop_validation.second_hidden_activation_node.fProp_matrix, scores);
stop_timer(4);
// And softmax and loss. Be careful of short minibatch
double minibatch_log_likelihood;
start_timer(5);
SoftmaxLogLoss().fProp(scores.leftCols(current_minibatch_size),
minibatch.row(ngram_size-1),
output_probs,
minibatch_log_likelihood);
stop_timer(5);
log_likelihood += minibatch_log_likelihood;
}
cerr << "Validation log-likelihood: "<< log_likelihood << '\n';
cerr << " perplexity: "<< exp(-log_likelihood/validation_data_size) << '\n';
// If the validation perplexity decreases, halve the learning rate.
if (current_validation_ll != 0.0 && log_likelihood < current_validation_ll && myParam.parameter_update != "ADA")
{
current_learning_rate /= 2;
}
current_validation_ll = log_likelihood;
}