if cfg.save_best:
global_valid = valid_metric.results_class_wise_average_metrics(
)['f_measure']['f_measure']
if not no_weak:
global_valid += np.mean(weak_metric)
if save_best_cb.apply(global_valid):
model_fname = os.path.join(saved_model_dir, "baseline_best")
torch.save(state, model_fname)
if cfg.save_best:
model_fname = os.path.join(saved_model_dir, "baseline_best")
state = torch.load(model_fname)
LOG.info("testing model: {}".format(model_fname))
else:
LOG.info("testing model of last epoch: {}".format(cfg.n_epoch))
# ##############
# Validation
# ##############
predicitons_fname = os.path.join(saved_pred_dir, "baseline_validation.tsv")
test_model(state, cfg.validation, reduced_number_of_data,
predicitons_fname)
# ##############
# Evaluation
# ##############
predicitons_eval2019_fname = os.path.join(saved_pred_dir,
"baseline_eval2019.tsv")
test_model(state, cfg.eval_desed, reduced_number_of_data,
predicitons_eval2019_fname)
np.mean(weak_metric)))
state['model']['state_dict'] = crnn.state_dict()
state['model_ema']['state_dict'] = crnn_ema.state_dict()
state['optimizer']['state_dict'] = optimizer.state_dict()
state['epoch'] = epoch
# state['valid_metric'] = valid_events_metric.results()
if cfg.checkpoint_epochs is not None and (
epoch + 1) % cfg.checkpoint_epochs == 0:
model_fname = os.path.join(saved_model_dir, '_epoch_' + str(epoch))
torch.save(state, model_fname)
if cfg.save_best:
global_valid = np.mean(weak_metric)
if save_best_cb.apply(global_valid):
model_fname = os.path.join(saved_model_dir, '_best')
torch.save(state, model_fname)
if cfg.save_best:
model_fname = os.path.join(saved_model_dir, '_best')
state = torch.load(model_fname)
LOG.info("testing model: {}".format(model_fname))
else:
LOG.info("testing model of last epoch: {}".format(cfg.n_epoch))
# ##############
# Validation
# ##############
predictions_fname = os.path.join(saved_pred_dir, "_validation.csv")
test_model(state, reduced_number_of_data, predictions_fname)