def get_SELD_Results(self, pred_files_path):
# collect predicted files info
pred_files = os.listdir(pred_files_path)
eval = SELD_evaluation_metrics.SELDMetrics(nb_classes=self._feat_cls.get_nb_classes(), doa_threshold=self._doa_thresh)
for pred_cnt, pred_file in enumerate(pred_files):
# Load predicted output format file
pred_dict = self._feat_cls.load_output_format_file(os.path.join(pred_files_path, pred_file))
if self._use_polar_format:
pred_dict = self._feat_cls.convert_output_format_cartesian_to_polar(pred_dict)
pred_labels = self._feat_cls.segment_labels(pred_dict, self._feat_cls.get_nb_frames())
# Calculated scores
eval.update_seld_scores(pred_labels, self._ref_labels[pred_file])
# Overall SED and DOA scores
ER, F, LE, LR = eval.compute_seld_scores()
seld_scr = SELD_evaluation_metrics.early_stopping_metric([ER, F], [LE, LR])
return ER, F, LE, LR, seld_scr
os.path.join(ref_desc_files, pred_file.replace('.npy',
'.csv')))
if use_polar_format:
gt_labels = feat_cls.segment_labels(gt_dict_polar,
feat_cls.get_nb_frames())
else:
gt_dict = feat_cls.convert_output_format_polar_to_cartesian(
gt_dict_polar)
gt_labels = feat_cls.segment_labels(gt_dict,
feat_cls.get_nb_frames())
# Calculated scores
if use_polar_format:
eval.update_seld_scores(pred_labels, gt_labels)
else:
eval.update_seld_scores_xyz(pred_labels, gt_labels)
# Overall SED and DOA scores
er, f, de, de_f = eval.compute_seld_scores()
seld_scr = SELD_evaluation_metrics.early_stopping_metric([er, f],
[de, de_f])
print('\nAverage score for {} {} data using {} coordinates'.format(
score_type, 'fold' if score_type == 'all' else split_key,
'Polar' if use_polar_format else 'Cartesian'))
print('SELD score (early stopping metric): {:0.2f}'.format(seld_scr))
print('SED metrics: Error rate: {:0.2f}, F-score:{:0.1f}'.format(
er, 100 * f))
print('DOA metrics: DOA error: {:0.1f}, F-score:{:0.1f}'.format(
de, 100 * de_f))
def get_consolidated_SELD_results(self,
pred_files_path,
score_type_list=['all', 'room']):
'''
Get all categories of results.
;score_type_list: Supported
'all' - all the predicted files
'room' - for individual rooms
'''
# collect predicted files info
pred_files = os.listdir(pred_files_path)
nb_pred_files = len(pred_files)
# Calculate scores for different splits, overlapping sound events, and impulse responses (reverberant scenes)
print('Number of predicted files: {}\nNumber of reference files: {}'.
format(nb_pred_files, self._nb_ref_files))
print('\nCalculating {} scores for {}'.format(
score_type_list, os.path.basename(pred_output_format_files)))
for score_type in score_type_list:
print(
'\n\n---------------------------------------------------------------------------------------------------'
)
print(
'------------------------------------ {} ---------------------------------------------'
.format('Total score' if score_type ==
'all' else 'score per {}'.format(score_type)))
print(
'---------------------------------------------------------------------------------------------------'
)
split_cnt_dict = self.get_nb_files(
pred_files,
tag=score_type) # collect files corresponding to score_type
# Calculate scores across files for a given score_type
for split_key in np.sort(list(split_cnt_dict)):
# Load evaluation metric class
eval = SELD_evaluation_metrics.SELDMetrics(
nb_classes=self._feat_cls.get_nb_classes(),
doa_threshold=self._doa_thresh)
for pred_cnt, pred_file in enumerate(
split_cnt_dict[split_key]):
# Load predicted output format file
pred_dict = self._feat_cls.load_output_format_file(
os.path.join(pred_output_format_files, pred_file))
if self._use_polar_format:
pred_dict = self._feat_cls.convert_output_format_cartesian_to_polar(
pred_dict)
pred_labels = self._feat_cls.segment_labels(
pred_dict, self._feat_cls.get_nb_frames())
# Calculated scores
eval.update_seld_scores(pred_labels,
self._ref_labels[pred_file])
# Overall SED and DOA scores
ER, F, LE, LR = eval.compute_seld_scores()
seld_scr = SELD_evaluation_metrics.early_stopping_metric(
[ER, F], [LE, LR])
print('\nAverage score for {} {} data using {} coordinates'.
format(
score_type,
'fold' if score_type == 'all' else split_key,
'Polar' if self._use_polar_format else 'Cartesian'))
print('SELD score (early stopping metric): {:0.2f}'.format(
seld_scr))
print(
'SED metrics: Error rate: {:0.2f}, F-score:{:0.1f}'.format(
ER, 100 * F))
print(
'DOA metrics: Localization error: {:0.1f}, Localization Recall: {:0.1f}'
.format(LE, 100 * LR))