def getTrainingFilenames(data_path):
"""
training filenames
:param data_path:
:return:
"""
fns = getRecordings(data_path)
train_fns = [fn.replace('feature_', '') for fn in fns if 'feature' in fn]
# train_fns = [fn.replace(fn.split('_')[0]+'_', '') for fn in fns]
return train_fns
def getTrainingFilenames(annotation_path, cv_filename):
"""
annotation filenames - cv test filenames
:param annotation_path:
:param cv_filename:
:return:
"""
annotation_fns = getRecordings(annotation_path)
test_fns = annotationCvParser(cv_filename)
train_fns = [x for x in annotation_fns if x not in test_fns]
return train_fns
def dump_feature_label_sample_weights_onset_phrase_bock(
audio_path, annotation_path, path_output):
"""
dump feature, label, sample weights for each phrase with bock annotation format
:param audio_path:
:param annotation_path:
:param path_output:
:return:
"""
log_mel_line_all = []
for fn in getRecordings(annotation_path):
# from the annotation to get feature, frame start and frame end of each line, frames_onset
log_mel, frames_onset, frame_start, frame_end = dump_feature_onset_helper(
audio_path, annotation_path, fn, 1)
# simple sample weighting
feature, label, sample_weights = \
feature_onset_phrase_label_sample_weights(frames_onset,
frame_start,
frame_end,
log_mel)
# save feature, label and weights
feature_label_weights_saver(path_output, fn, feature, label,
sample_weights)
log_mel_line_all.append(feature)
log_mel_line_all = np.concatenate(log_mel_line_all)
scaler = preprocessing.StandardScaler()
scaler.fit(log_mel_line_all)
pickle.dump(scaler, open(path_output, 'wb'))
return True
def save_feature_label_sample_weights_onset_phrase(wav_path, textgrid_path,
path_output, tier_parent,
tier_child):
"""
phrase level feature and sample weights dump
:param wav_path:
:param textgrid_path:
:param recordings:
:param tier_parent:
:param tier_child
:return:
"""
recordings = getRecordings(textgrid_path)
log_mel_line_all = []
for recording_name in recordings:
nested_utterance_lists, log_mel = dump_training_data_textgrid_helper(
wav_path=wav_path,
textgrid_path=textgrid_path,
recording_name=recording_name,
tier_parent=tier_parent,
tier_child=tier_child)
# create the ground truth lab files
for idx, u_list in enumerate(nested_utterance_lists):
print 'Processing feature collecting ... ' + recording_name + ' phrase ' + str(
idx + 1)
if not len(u_list[1]):
continue
frames_onset, frame_start, frame_end = \
get_onset_in_frame_helper(recording_name, idx, lab=False, u_list=u_list)
log_mel_line, label, sample_weights = \
feature_onset_phrase_label_sample_weights(frames_onset, frame_start, frame_end, log_mel)
# save feature in h5py
filename_log_mel_line = join(
path_output,
'feature' + '_' + recording_name + '_' + str(idx) + '.h5')
h5f = h5py.File(filename_log_mel_line, 'w')
h5f.create_dataset('feature_all', data=log_mel_line)
h5f.close()
# dumpy label
filename_label = join(
path_output,
'label' + '_' + recording_name + '_' + str(idx) + '.pkl')
pickle.dump(label, open(filename_label, 'wb'), protocol=2)
# dump sample weights
filename_sample_weights = join(
path_output, 'sample_weights' + '_' + recording_name + '_' +
str(idx) + '.pkl')
pickle.dump(sample_weights,
open(filename_sample_weights, 'wb'),
protocol=2)
log_mel_line_all.append(log_mel_line)
log_mel_line_all = np.concatenate(log_mel_line_all)
# dump scaler
scaler = preprocessing.StandardScaler()
scaler.fit(log_mel_line_all)
filename_scaler = join(path_output, 'scaler_phrase.pkl')
pickle.dump(scaler, open(filename_scaler, 'wb'))
return True
def dump_feature_label_sample_weights(path_audio,
path_textgrid,
path_output=None,
tier_parent=None,
tier_child=None):
"""
dump features for all the dataset for onset detection
:return:
"""
recording_names = getRecordings(path_textgrid)
log_mel_p, \
log_mel_n, \
sample_weights_p, \
sample_weights_n \
= dump_feature_sample_weights_onset(wav_path=path_audio,
textgrid_path=path_textgrid,
recordings=recording_names,
tier_parent=tier_parent,
tier_child=tier_child)
print('finished feature extraction.')
sample_weights = np.concatenate((sample_weights_p, sample_weights_n))
# save h5py separately for postive and negative features
filename_log_mel_p = join(path_output, 'log_mel_p.h5')
h5f = h5py.File(filename_log_mel_p, 'w')
h5f.create_dataset('mfcc_p', data=log_mel_p)
h5f.close()
filename_log_mel_n = join(feature_data_path, 'log_mel_n.h5')
h5f = h5py.File(filename_log_mel_n, 'w')
h5f.create_dataset('mfcc_n', data=log_mel_n)
h5f.close()
del log_mel_p
del log_mel_n
feature_all, label_all, scaler = \
feature_label_concatenation_h5py(filename_log_mel_p,
filename_log_mel_n,
scaling=True)
print('finished feature concatenation.')
os.remove(filename_log_mel_p)
os.remove(filename_log_mel_n)
feature_all = featureReshape(feature_all, nlen=7)
print('feature shape:', feature_all.shape)
# save feature
filename_feature_all = join(path_output, 'feature.h5')
h5f = h5py.File(filename_feature_all, 'w')
h5f.create_dataset('feature_all', data=feature_all)
h5f.close()
# save label, sample weights and scaler
pickle.dump(label_all,
open(join(path_output, 'labels.pkl'), 'wb'),
protocol=2)
pickle.dump(sample_weights,
open(join(path_output, 'sample_weights.pkl'), 'wb'),
protocol=2)
pickle.dump(scaler,
open(join(path_output, 'scaler.pkl'), 'wb'),
protocol=2)
elif v.lower() in ('no', 'false', 'f', 'n', '0'):
return False
else:
raise argparse.ArgumentTypeError('Boolean value expected.')
parser = argparse.ArgumentParser(description="dump feature, label and sample weights for bock train set.")
parser.add_argument("-p",
"--phrase",
type=str2bool,
default='true',
help="whether to collect feature for each phrase")
args = parser.parse_args()
if args.phrase:
mfcc_line_all = []
for fn in getRecordings(bock_annotations_path):
mfcc_line = dump_feature_onset_phrase(audio_path=bock_audio_path,
annotation_path=bock_annotations_path,
fn=fn,
channel=1)
mfcc_line_all.append(mfcc_line)
mfcc_line_all = np.concatenate(mfcc_line_all)
scaler = preprocessing.StandardScaler()
scaler.fit(mfcc_line_all)
pickle.dump(scaler, open(join(bock_feature_data_path_madmom_simpleSampleWeighting_phrase,
'scaler_bock_phrase.pkl'), 'wb'))
# else: