Skip to content

Finish training if we cannot get better dev ppl. for N hours. #5

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
MorinoseiMorizo wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Finish training if we cannot get better dev ppl. for N hours. #5

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bf29322
add function to stop training when we cannot get lower dev ppl for so…
MorinoseiMorizo Jan 6, 2017
3d0b404
add number of trained sample to evaluation log.
MorinoseiMorizo Jan 6, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions sample_data/sample_config.ini
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,6 +180,10 @@ dropout_ratio=0.3
; Maximum number of batch data to be trained.
max_iteration=10000

; Training will be stopped if we cannot get the lowest dev ppl. for this hours.
; If this is set to 0, we don't stop training with this criteria.
max_waiting_hour=24
Copy link
Owner

@odashi odashi Jan 13, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this method could not stop the trainer as soon as the dev-loss goes worse. Maybe it is better to remove this behavior and to recommend to set some large value (e.g. 114514) to disable early stopping virtually.


; Timing of evaluating dev/test set. Available options:
; * step .... Number of steps (iterations).
; * sample .... Number of samples (sentences).
Expand Down
28 changes: 20 additions & 8 deletions src/bin/train.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -450,6 +450,10 @@ int main(int argc, char * argv[]) {
const float dropout_ratio = config.get<float>("Train.dropout_ratio");
const unsigned max_iteration = config.get<unsigned>("Train.max_iteration");

const unsigned max_waiting_hour = config.get<unsigned>("Train.max_waiting_hour");
auto scheduled_training_finish_time = std::chrono::system_clock::to_time_t(
std::chrono::system_clock::now() + std::chrono::hours(max_waiting_hour));

const string eval_type = config.get<string>("Train.evaluation_type");
const unsigned eval_interval = config.get<unsigned>(
"Train.evaluation_interval");
Expand Down Expand Up @@ -532,29 +536,29 @@ int main(int argc, char * argv[]) {
const float dev_log_ppl = ::evaluateLogPerplexity(
encdec, dev_sampler, batch_converter);
const auto fmt_dev_log_ppl = boost::format(
"Evaluated: batch=%d words=%d elapsed-time(sec)=%d dev-log-ppl=%.6e")
% iteration % num_trained_words % elapsed_time_seconds % dev_log_ppl;
"Evaluated: batch=%d samples=%d words=%d elapsed-time(sec)=%d dev-log-ppl=%.6e")
% iteration % num_trained_samples % num_trained_words % elapsed_time_seconds % dev_log_ppl;
logger->info(fmt_dev_log_ppl.str());

const float dev_bleu = ::evaluateBLEU(
*trg_vocab, encdec, dev_sampler, train_max_length);
const auto fmt_dev_bleu = boost::format(
"Evaluated: batch=%d words=%d elapsed-time(sec)=%d dev-bleu=%.6f")
% iteration % num_trained_words % elapsed_time_seconds % dev_bleu;
"Evaluated: batch=%d samples=%d words=%d elapsed-time(sec)=%d dev-bleu=%.6f")
% iteration % num_trained_samples % num_trained_words % elapsed_time_seconds % dev_bleu;
logger->info(fmt_dev_bleu.str());

const float test_log_ppl = ::evaluateLogPerplexity(
encdec, test_sampler, batch_converter);
const auto fmt_test_log_ppl = boost::format(
"Evaluated: batch=%d words=%d elapsed-time(sec)=%d test-log-ppl=%.6e")
% iteration % num_trained_words % elapsed_time_seconds % test_log_ppl;
"Evaluated: batch=%d samples=%d words=%d elapsed-time(sec)=%d test-log-ppl=%.6e")
% iteration % num_trained_samples % num_trained_words % elapsed_time_seconds % test_log_ppl;
logger->info(fmt_test_log_ppl.str());

const float test_bleu = ::evaluateBLEU(
*trg_vocab, encdec, test_sampler, train_max_length);
const auto fmt_test_bleu = boost::format(
"Evaluated: batch=%d words=%d elapsed-time(sec)=%d test-bleu=%.6f")
% iteration % num_trained_words % elapsed_time_seconds % test_bleu;
"Evaluated: batch=%d samples=%d words=%d elapsed-time(sec)=%d test-bleu=%.6f")
% iteration % num_trained_samples % num_trained_words % elapsed_time_seconds % test_bleu;
logger->info(fmt_test_bleu.str());

if (lr_decay_type == "eval") {
Expand All @@ -576,6 +580,8 @@ int main(int argc, char * argv[]) {
FS::copy_file(model_dir / "latest.trainer.params", trainer_path);
FS::copy_file(model_dir / "latest.model.params", model_path);
logger->info("Saved 'best_dev_log_ppl' model.");
scheduled_training_finish_time = std::chrono::system_clock::to_time_t(
std::chrono::system_clock::now() + std::chrono::hours(max_waiting_hour));
} else {
if (lr_decay_type == "logppl") {
lr_decay *= lr_decay_ratio;
Expand Down Expand Up @@ -608,6 +614,12 @@ int main(int argc, char * argv[]) {
std::chrono::system_clock::to_time_t(
current_time + std::chrono::minutes(eval_interval));
}

// Training finish check
if (max_waiting_hour != 0 and
std::chrono::system_clock::to_time_t(std::chrono::system_clock::now()) >= scheduled_training_finish_time) {
break;
}
}

logger->info("Finished.");
Expand Down