def metric_dcase2019(gen, sed_pred, doa_pred):
sed_gt = gen.all_label_sed_2d()
doa_gt = gen.all_label_doa_2d()
sed_metric = evaluation_metrics.compute_sed_scores(
sed_pred, sed_gt, feat_cls.nb_frames_1s())
doa_metric = evaluation_metrics.compute_doa_scores_regr_xyz(
doa_pred, doa_gt, sed_pred, sed_gt)
seld_metric = evaluation_metrics.early_stopping_metric(
sed_metric, doa_metric)
return sed_metric, doa_metric, seld_metric
def metric_dcase2020(gen, sed_pred, doa_pred):
sed_gt = gen.all_label_sed_2d()
doa_gt = gen.all_label_doa_2d()
cls_new_metric = SELD_evaluation_metrics.SELDMetrics(
nb_classes=gen._Ncat, doa_threshold=params['lad_doa_thresh'])
pred_dict = feat_cls.regression_label_format_to_output_format(
sed_pred, doa_pred)
gt_dict = feat_cls.regression_label_format_to_output_format(
sed_gt, doa_gt)
pred_blocks_dict = feat_cls.segment_labels(pred_dict,
sed_pred.shape[0])
gt_blocks_dict = feat_cls.segment_labels(gt_dict, sed_gt.shape[0])
cls_new_metric.update_seld_scores_xyz(pred_blocks_dict,
gt_blocks_dict)
new_metric = cls_new_metric.compute_seld_scores()
new_seld_metric = evaluation_metrics.early_stopping_metric(
new_metric[:2], new_metric[2:])
return new_metric, new_seld_metric
def main(argv):
"""
Main wrapper for training sound event localization and detection network.
:param argv: expects two optional inputs.
first input: task_id - (optional) To chose the system configuration in parameters.py.
(default) 1 - uses default parameters
second input: job_id - (optional) all the output files will be uniquely represented with this.
(default) 1
"""
print(argv)
if len(argv) != 3:
print('\n\n')
print(
'-------------------------------------------------------------------------------------------------------'
)
print('The code expected two optional inputs')
print('\t>> python seld.py <task-id> <job-id>')
print(
'\t\t<task-id> is used to choose the user-defined parameter set from parameter.py'
)
print('Using default inputs for now')
print(
'\t\t<job-id> is a unique identifier which is used for output filenames (models, training plots). '
'You can use any number or string for this.')
print(
'-------------------------------------------------------------------------------------------------------'
)
print('\n\n')
# use parameter set defined by user
task_id = '1' if len(argv) < 2 else argv[1]
params = parameter.get_params(task_id)
job_id = 1 if len(argv) < 3 else argv[-1]
feat_cls = cls_feature_class.FeatureClass(params)
train_splits, val_splits, test_splits = None, None, None
if params['mode'] == 'dev':
test_splits = [1]
val_splits = [2]
train_splits = [[3, 4, 5, 6]]
elif params['mode'] == 'eval':
test_splits = [[7, 8]]
val_splits = [[1]]
train_splits = [[2, 3, 4, 5, 6]]
avg_scores_val = []
avg_scores_test = []
for split_cnt, split in enumerate(test_splits):
print(
'\n\n---------------------------------------------------------------------------------------------------'
)
print(
'------------------------------------ SPLIT {} -----------------------------------------------'
.format(split))
print(
'---------------------------------------------------------------------------------------------------'
)
# Unique name for the run
cls_feature_class.create_folder(params['model_dir'])
unique_name = '{}_{}_{}_{}_split{}'.format(task_id, job_id,
params['dataset'],
params['mode'], split)
unique_name = os.path.join(params['model_dir'], unique_name)
model_name = '{}_model.h5'.format(unique_name)
print("unique_name: {}\n".format(unique_name))
# Load train and validation data
print('Loading training dataset:')
data_gen_train = cls_data_generator.DataGenerator(
params=params, split=train_splits[split_cnt])
print('Loading validation dataset:')
data_gen_val = cls_data_generator.DataGenerator(
params=params, split=val_splits[split_cnt], shuffle=False)
# Collect the reference labels for validation data
data_in, data_out = data_gen_train.get_data_sizes()
print('FEATURES:\n\tdata_in: {}\n\tdata_out: {}\n'.format(
data_in, data_out))
nb_classes = data_gen_train.get_nb_classes()
gt = collect_test_labels(data_gen_val, data_out, nb_classes,
params['quick_test'])
sed_gt = evaluation_metrics.reshape_3Dto2D(gt[0])
doa_gt = evaluation_metrics.reshape_3Dto2D(gt[1])
print(
'MODEL:\n\tdropout_rate: {}\n\tCNN: nb_cnn_filt: {}, f_pool_size{}, t_pool_size{}\n\trnn_size: {}, fnn_size: {}\n\tdoa_objective: {}\n'
.format(params['dropout_rate'], params['nb_cnn2d_filt'],
params['f_pool_size'], params['t_pool_size'],
params['rnn_size'], params['fnn_size'],
params['doa_objective']))
print('Using loss weights : {}'.format(params['loss_weights']))
model = keras_model.get_model(data_in=data_in,
data_out=data_out,
dropout_rate=params['dropout_rate'],
nb_cnn2d_filt=params['nb_cnn2d_filt'],
f_pool_size=params['f_pool_size'],
t_pool_size=params['t_pool_size'],
rnn_size=params['rnn_size'],
fnn_size=params['fnn_size'],
weights=params['loss_weights'],
doa_objective=params['doa_objective'])
best_seld_metric = 99999
best_epoch = -1
patience_cnt = 0
nb_epoch = 2 if params['quick_test'] else params['nb_epochs']
seld_metric = np.zeros(nb_epoch)
new_seld_metric = np.zeros(nb_epoch)
tr_loss = np.zeros(nb_epoch)
doa_metric = np.zeros((nb_epoch, 6))
sed_metric = np.zeros((nb_epoch, 2))
new_metric = np.zeros((nb_epoch, 4))
# start training
for epoch_cnt in range(nb_epoch):
start = time.time()
# train once per epoch
hist = model.fit_generator(
generator=data_gen_train.generate(),
steps_per_epoch=2 if params['quick_test'] else
data_gen_train.get_total_batches_in_data(),
epochs=params['epochs_per_fit'],
verbose=2,
)
tr_loss[epoch_cnt] = hist.history.get('loss')[-1]
# predict once per peoch
pred = model.predict_generator(
generator=data_gen_val.generate(),
steps=2 if params['quick_test'] else
data_gen_val.get_total_batches_in_data(),
verbose=2)
sed_pred = evaluation_metrics.reshape_3Dto2D(pred[0]) > 0.5
doa_pred = evaluation_metrics.reshape_3Dto2D(
pred[1]
if params['doa_objective'] is 'mse' else pred[1][:, :,
nb_classes:])
# Calculate the DCASE 2019 metrics - Detection-only and Localization-only scores
sed_metric[epoch_cnt, :] = evaluation_metrics.compute_sed_scores(
sed_pred, sed_gt, data_gen_val.nb_frames_1s())
doa_metric[
epoch_cnt, :] = evaluation_metrics.compute_doa_scores_regr_xyz(
doa_pred, doa_gt, sed_pred, sed_gt)
seld_metric[epoch_cnt] = evaluation_metrics.early_stopping_metric(
sed_metric[epoch_cnt, :], doa_metric[epoch_cnt, :])
# Calculate the DCASE 2020 metrics - Location-aware detection and Class-aware localization scores
cls_new_metric = SELD_evaluation_metrics.SELDMetrics(
nb_classes=data_gen_val.get_nb_classes(),
doa_threshold=params['lad_doa_thresh'])
pred_dict = feat_cls.regression_label_format_to_output_format(
sed_pred, doa_pred)
gt_dict = feat_cls.regression_label_format_to_output_format(
sed_gt, doa_gt)
pred_blocks_dict = feat_cls.segment_labels(pred_dict,
sed_pred.shape[0])
gt_blocks_dict = feat_cls.segment_labels(gt_dict, sed_gt.shape[0])
cls_new_metric.update_seld_scores_xyz(pred_blocks_dict,
gt_blocks_dict)
new_metric[epoch_cnt, :] = cls_new_metric.compute_seld_scores()
new_seld_metric[
epoch_cnt] = evaluation_metrics.early_stopping_metric(
new_metric[epoch_cnt, :2], new_metric[epoch_cnt, 2:])
# Visualize the metrics with respect to epochs
plot_functions(unique_name, tr_loss, sed_metric, doa_metric,
seld_metric, new_metric, new_seld_metric)
patience_cnt += 1
if new_seld_metric[epoch_cnt] < best_seld_metric:
best_seld_metric = new_seld_metric[epoch_cnt]
best_epoch = epoch_cnt
model.save(model_name)
patience_cnt = 0
print(
'epoch_cnt: {}, time: {:0.2f}s, tr_loss: {:0.2f}, '
'\n\t\t DCASE2019 SCORES: ER: {:0.2f}, F: {:0.1f}, DE: {:0.1f}, FR:{:0.1f}, seld_score: {:0.2f}, '
'\n\t\t DCASE2020 SCORES: ER: {:0.2f}, F: {:0.1f}, DE: {:0.1f}, DE_F:{:0.1f}, seld_score (early stopping score): {:0.2f}, '
'best_seld_score: {:0.2f}, best_epoch : {}\n'.format(
epoch_cnt,
time.time() - start, tr_loss[epoch_cnt],
sed_metric[epoch_cnt, 0], sed_metric[epoch_cnt, 1] * 100,
doa_metric[epoch_cnt, 0], doa_metric[epoch_cnt, 1] * 100,
seld_metric[epoch_cnt], new_metric[epoch_cnt, 0],
new_metric[epoch_cnt, 1] * 100, new_metric[epoch_cnt, 2],
new_metric[epoch_cnt, 3] * 100, new_seld_metric[epoch_cnt],
best_seld_metric, best_epoch))
if patience_cnt > params['patience']:
break
avg_scores_val.append([
new_metric[best_epoch, 0], new_metric[best_epoch, 1],
new_metric[best_epoch, 2], new_metric[best_epoch,
3], best_seld_metric
])
print('\nResults on validation split:')
print('\tUnique_name: {} '.format(unique_name))
print('\tSaved model for the best_epoch: {}'.format(best_epoch))
print('\tSELD_score (early stopping score) : {}'.format(
best_seld_metric))
print('\n\tDCASE2020 scores')
print(
'\tClass-aware localization scores: DOA_error: {:0.1f}, F-score: {:0.1f}'
.format(new_metric[best_epoch, 2],
new_metric[best_epoch, 3] * 100))
print(
'\tLocation-aware detection scores: Error rate: {:0.2f}, F-score: {:0.1f}'
.format(new_metric[best_epoch, 0],
new_metric[best_epoch, 1] * 100))
print('\n\tDCASE2019 scores')
print(
'\tLocalization-only scores: DOA_error: {:0.1f}, Frame recall: {:0.1f}'
.format(doa_metric[best_epoch, 0],
doa_metric[best_epoch, 1] * 100))
print(
'\tDetection-only scores: Error rate: {:0.2f}, F-score: {:0.1f}\n'.
format(sed_metric[best_epoch, 0], sed_metric[best_epoch, 1] * 100))
# ------------------ Calculate metric scores for unseen test split ---------------------------------
print(
'\nLoading the best model and predicting results on the testing split'
)
print('\tLoading testing dataset:')
data_gen_test = cls_data_generator.DataGenerator(
params=params,
split=split,
shuffle=False,
per_file=params['dcase_output'],
is_eval=True if params['mode'] is 'eval' else False)
model = keras_model.load_seld_model('{}_model.h5'.format(unique_name),
params['doa_objective'])
pred_test = model.predict_generator(
generator=data_gen_test.generate(),
steps=2 if params['quick_test'] else
data_gen_test.get_total_batches_in_data(),
verbose=2)
test_sed_pred = evaluation_metrics.reshape_3Dto2D(pred_test[0]) > 0.5
test_doa_pred = evaluation_metrics.reshape_3Dto2D(
pred_test[1]
if params['doa_objective'] is 'mse' else pred_test[1][:, :,
nb_classes:])
if params['dcase_output']:
# Dump results in DCASE output format for calculating final scores
dcase_dump_folder = os.path.join(
params['dcase_dir'],
'{}_{}_{}'.format(task_id, params['dataset'], params['mode']))
cls_feature_class.create_folder(dcase_dump_folder)
print('Dumping recording-wise results in: {}'.format(
dcase_dump_folder))
test_filelist = data_gen_test.get_filelist()
# Number of frames for a 60 second audio with 20ms hop length = 3000 frames
max_frames_with_content = data_gen_test.get_nb_frames()
# Number of frames in one batch (batch_size* sequence_length) consists of all the 3000 frames above with
# zero padding in the remaining frames
frames_per_file = data_gen_test.get_frame_per_file()
for file_cnt in range(test_sed_pred.shape[0] // frames_per_file):
output_file = os.path.join(
dcase_dump_folder,
test_filelist[file_cnt].replace('.npy', '.csv'))
dc = file_cnt * frames_per_file
output_dict = feat_cls.regression_label_format_to_output_format(
test_sed_pred[dc:dc + max_frames_with_content, :],
test_doa_pred[dc:dc + max_frames_with_content, :])
data_gen_test.write_output_format_file(output_file,
output_dict)
if params['mode'] is 'dev':
test_data_in, test_data_out = data_gen_test.get_data_sizes()
test_gt = collect_test_labels(data_gen_test, test_data_out,
nb_classes, params['quick_test'])
test_sed_gt = evaluation_metrics.reshape_3Dto2D(test_gt[0])
test_doa_gt = evaluation_metrics.reshape_3Dto2D(test_gt[1])
# Calculate DCASE2019 scores
test_sed_loss = evaluation_metrics.compute_sed_scores(
test_sed_pred, test_sed_gt, data_gen_test.nb_frames_1s())
test_doa_loss = evaluation_metrics.compute_doa_scores_regr_xyz(
test_doa_pred, test_doa_gt, test_sed_pred, test_sed_gt)
test_metric_loss = evaluation_metrics.early_stopping_metric(
test_sed_loss, test_doa_loss)
# Calculate DCASE2020 scores
cls_new_metric = SELD_evaluation_metrics.SELDMetrics(
nb_classes=data_gen_test.get_nb_classes(), doa_threshold=20)
test_pred_dict = feat_cls.regression_label_format_to_output_format(
test_sed_pred, test_doa_pred)
test_gt_dict = feat_cls.regression_label_format_to_output_format(
test_sed_gt, test_doa_gt)
test_pred_blocks_dict = feat_cls.segment_labels(
test_pred_dict, test_sed_pred.shape[0])
test_gt_blocks_dict = feat_cls.segment_labels(
test_gt_dict, test_sed_gt.shape[0])
cls_new_metric.update_seld_scores_xyz(test_pred_blocks_dict,
test_gt_blocks_dict)
test_new_metric = cls_new_metric.compute_seld_scores()
test_new_seld_metric = evaluation_metrics.early_stopping_metric(
test_new_metric[:2], test_new_metric[2:])
avg_scores_test.append([
test_new_metric[0], test_new_metric[1], test_new_metric[2],
test_new_metric[3], test_new_seld_metric
])
print('Results on test split:')
print('\tDCASE2020 Scores')
print(
'\tClass-aware localization scores: DOA Error: {:0.1f}, F-score: {:0.1f}'
.format(test_new_metric[2], test_new_metric[3] * 100))
print(
'\tLocation-aware detection scores: Error rate: {:0.2f}, F-score: {:0.1f}'
.format(test_new_metric[0], test_new_metric[1] * 100))
print('\tSELD (early stopping metric): {:0.2f}'.format(
test_new_seld_metric))
print('\n\tDCASE2019 Scores')
print(
'\tLocalization-only scores: DOA Error: {:0.1f}, Frame recall: {:0.1f}'
.format(test_doa_loss[0], test_doa_loss[1] * 100))
print(
'\tDetection-only scores:Error rate: {:0.2f}, F-score: {:0.1f}'
.format(test_sed_loss[0], test_sed_loss[1] * 100))