def test_file_wise_generator(PARAMS, file_name_sp, file_name_mu, target_dB):
n_fft = PARAMS['n_fft'][PARAMS['Model']]
n_mels = PARAMS['n_mels'][PARAMS['Model']]
featName = PARAMS['featName'][PARAMS['Model']]
if file_name_mu=='':
fv = preproc.get_featuregram(PARAMS, 'speech', PARAMS['feature_opDir'], file_name_sp, '', None, n_fft, n_mels, featName, save_feat=True)
elif file_name_sp=='':
fv = preproc.get_featuregram(PARAMS, 'music', PARAMS['feature_opDir'], '', file_name_mu, None, n_fft, n_mels, featName, save_feat=True)
else:
fv = preproc.get_featuregram(PARAMS, 'speech_music', PARAMS['feature_opDir'], file_name_sp, file_name_mu, target_dB, n_fft, n_mels, featName, save_feat=True)
if PARAMS['frame_level_scaling']:
fv = cscale_data(fv, PARAMS['mean_fold'+str(PARAMS['fold'])], PARAMS['stdev_fold'+str(PARAMS['fold'])])
batchData = preproc.get_feature_patches(PARAMS, fv, PARAMS['W'], PARAMS['W_shift'], featName)
if PARAMS['skewness_vector']:
if PARAMS['skewness_vector']=='Harm':
batchData = cget_data_statistics(batchData, axis=1)
batchData = np.expand_dims(batchData, axis=2)
elif PARAMS['skewness_vector']=='Perc':
batchData = cget_data_statistics(batchData, axis=0)
batchData = np.expand_dims(batchData, axis=1)
if 'Lemaire_et_al' in PARAMS['Model']:
batchData = np.transpose(batchData, axes=(0,2,1)) # TCN input shape=(batch_size, timesteps, ndim)
numLab = np.shape(batchData)[0]
if file_name_mu=='': # speech
batchLabel = np.array([1]*numLab)
elif file_name_sp=='': # music
batchLabel = np.array([0]*numLab)
else: # speech+music
batchLabel = np.array([2]*numLab)
OHE_batchLabel = to_categorical(batchLabel, num_classes=len(PARAMS['classes']))
return batchData, OHE_batchLabel
def generator(PARAMS, folder, file_list, batchSize):
batch_count = 0
np.random.shuffle(file_list['speech'])
np.random.shuffle(file_list['music'])
file_list_sp_temp = file_list['speech'].copy()
file_list_mu_temp = file_list['music'].copy()
batchData_sp = np.empty([], dtype=float)
batchData_mu = np.empty([], dtype=float)
balance_sp = 0
balance_mu = 0
if not os.path.exists(PARAMS['feature_opDir'] + '/speech/'):
os.makedirs(PARAMS['feature_opDir'] + '/speech/')
if not os.path.exists(PARAMS['feature_opDir'] + '/music/'):
os.makedirs(PARAMS['feature_opDir'] + '/music/')
if len(PARAMS['classes']) == 3:
np.random.shuffle(file_list['speech+music'])
file_list_spmu_temp = file_list['speech+music'].copy()
batchData_spmu = np.empty([], dtype=float)
balance_spmu = 0
if not os.path.exists(PARAMS['feature_opDir'] + '/speech_music/'):
os.makedirs(PARAMS['feature_opDir'] + '/speech_music/')
n_fft = PARAMS['n_fft'][PARAMS['Model']]
n_mels = PARAMS['n_mels'][PARAMS['Model']]
featName = PARAMS['featName'][PARAMS['Model']]
while 1:
batchData = np.empty([], dtype=float)
while balance_sp < batchSize:
if not file_list_sp_temp:
file_list_sp_temp = file_list['speech'].copy()
sp_fName = file_list_sp_temp.pop()
sp_fName_path = folder + '/speech/' + sp_fName
if not os.path.exists(sp_fName_path):
# print(sp_fName_path, os.path.exists(sp_fName_path))
continue
fv_sp = preproc.get_featuregram(PARAMS, 'speech',
PARAMS['feature_opDir'],
sp_fName_path, '', None, n_fft,
n_mels, featName)
if PARAMS['frame_level_scaling']:
fv_sp = cscale_data(fv_sp,
PARAMS['mean_fold' + str(PARAMS['fold'])],
PARAMS['stdev_fold' + str(PARAMS['fold'])])
fv_sp_patches = preproc.get_feature_patches(
PARAMS, fv_sp, PARAMS['W'], PARAMS['W_shift'], featName)
if balance_sp == 0:
batchData_sp = fv_sp_patches
else:
batchData_sp = np.append(batchData_sp, fv_sp_patches, axis=0)
balance_sp += np.shape(fv_sp_patches)[0]
# print('Speech: ', batchSize, balance_sp, np.shape(batchData_sp))
while balance_mu < batchSize:
if not file_list_mu_temp:
file_list_mu_temp = file_list['music'].copy()
mu_fName = file_list_mu_temp.pop()
mu_fName_path = folder + '/music/' + mu_fName
if not os.path.exists(mu_fName_path):
# print(mu_fName_path, os.path.exists(mu_fName_path))
continue
fv_mu = preproc.get_featuregram(PARAMS, 'music',
PARAMS['feature_opDir'], '',
mu_fName_path, None, n_fft, n_mels,
featName)
if PARAMS['frame_level_scaling']:
fv_mu = cscale_data(fv_mu,
PARAMS['mean_fold' + str(PARAMS['fold'])],
PARAMS['stdev_fold' + str(PARAMS['fold'])])
fv_mu_patches = preproc.get_feature_patches(
PARAMS, fv_mu, PARAMS['W'], PARAMS['W_shift'], featName)
if balance_mu == 0:
batchData_mu = fv_mu_patches
else:
batchData_mu = np.append(batchData_mu, fv_mu_patches, axis=0)
balance_mu += np.shape(fv_mu_patches)[0]
# print('Music: ', batchSize, balance_mu, np.shape(batchData_mu))
batchData = batchData_mu[:batchSize, :] # music label=0
batchData = np.append(batchData, batchData_sp[:batchSize, :],
axis=0) # speech label=1
balance_mu -= batchSize
balance_sp -= batchSize
batchData_mu = batchData_mu[batchSize:, :]
batchData_sp = batchData_sp[batchSize:, :]
batchLabel = [0] * batchSize # music
batchLabel.extend([1] * batchSize) # speech
if len(PARAMS['classes']) == 3:
while balance_spmu < batchSize:
if not file_list_spmu_temp:
file_list_spmu_temp = file_list['speech+music'].copy()
np.random.shuffle(file_list_spmu_temp)
spmu_info = file_list_spmu_temp.pop()
sp_fName = spmu_info['speech']
sp_fName_path = folder + '/speech/' + sp_fName
mu_fName = spmu_info['music']
mu_fName_path = folder + '/music/' + mu_fName
target_dB = spmu_info['SMR']
if (not os.path.exists(mu_fName_path)) or (
not os.path.exists(sp_fName_path)):
continue
fv_spmu = preproc.get_featuregram(PARAMS, 'speech_music',
PARAMS['feature_opDir'],
sp_fName_path, mu_fName_path,
target_dB, n_fft, n_mels,
featName)
if PARAMS['frame_level_scaling']:
fv_spmu = cscale_data(
fv_spmu, PARAMS['mean_fold' + str(PARAMS['fold'])],
PARAMS['stdev_fold' + str(PARAMS['fold'])])
fv_spmu_patches = preproc.get_feature_patches(
PARAMS, fv_spmu, PARAMS['W'], PARAMS['W_shift'], featName)
if balance_spmu == 0:
batchData_spmu = fv_spmu_patches
else:
batchData_spmu = np.append(batchData_spmu,
fv_spmu_patches,
axis=0)
balance_spmu += np.shape(fv_spmu_patches)[0]
# print('SpeechMusic: ', batchSize, balance_spmu, np.shape(batchData_spmu))
# speech_music label=2
batchData = np.append(batchData,
batchData_spmu[:batchSize, :],
axis=0)
balance_spmu -= batchSize
batchData_spmu = batchData_spmu[batchSize:, :]
batchLabel.extend([2] * batchSize) # speech+music
if PARAMS['Model'] == 'Lemaire_et_al':
batchData = np.transpose(
batchData,
axes=(0, 2,
1)) # TCN input shape=(batch_size, timesteps, ndim)
''' Adding Normal (Gaussian) noise for data augmentation '''
if PARAMS['data_augmentation_with_noise']:
scale = np.random.choice([5e-3, 1e-3, 5e-4, 1e-4])
noise = np.random.normal(loc=0.0,
scale=scale,
size=np.shape(batchData))
batchData = np.add(batchData, noise)
OHE_batchLabel = to_categorical(batchLabel,
num_classes=len(PARAMS['classes']))
batch_count += 1
# print('Batch ', batch_count, ' shape=', np.shape(batchData), np.shape(OHE_batchLabel))
yield batchData, OHE_batchLabel
def generator(PARAMS, folder, file_list, batchSize):
batch_count = 0
np.random.shuffle(file_list['speech'])
np.random.shuffle(file_list['music'])
file_list_sp_temp = file_list['speech'].copy()
file_list_mu_temp = file_list['music'].copy()
batchData_sp = np.empty([], dtype=float)
batchData_mu = np.empty([], dtype=float)
balance_sp = 0
balance_mu = 0
if not os.path.exists(PARAMS['feature_opDir']+'/speech/'):
os.makedirs(PARAMS['feature_opDir']+'/speech/')
if not os.path.exists(PARAMS['feature_opDir']+'/music/'):
os.makedirs(PARAMS['feature_opDir']+'/music/')
if len(PARAMS['classes'])==3:
np.random.shuffle(file_list['speech+music'])
file_list_spmu_temp = file_list['speech+music'].copy()
batchData_spmu = np.empty([], dtype=float)
batchData_spmu_target_dB = np.empty([], dtype=float)
balance_spmu = 0
if not os.path.exists(PARAMS['feature_opDir']+'/speech_music/'):
os.makedirs(PARAMS['feature_opDir']+'/speech_music/')
n_fft = PARAMS['n_fft'][PARAMS['Model']]
n_mels = PARAMS['n_mels'][PARAMS['Model']]
featName = PARAMS['featName'][PARAMS['Model']]
while 1:
batchData = np.empty([], dtype=float)
while balance_sp<batchSize:
if not file_list_sp_temp:
file_list_sp_temp = file_list['speech'].copy()
sp_fName = file_list_sp_temp.pop()
sp_fName_path = folder + '/speech/' + sp_fName
if not os.path.exists(sp_fName_path):
# print(sp_fName_path, os.path.exists(sp_fName_path))
continue
fv_sp = preproc.get_featuregram(PARAMS, 'speech', PARAMS['feature_opDir'], sp_fName_path, '', None, n_fft, n_mels, featName)
if PARAMS['frame_level_scaling']:
fv_sp = cscale_data(fv_sp, PARAMS['mean_fold'+str(PARAMS['fold'])], PARAMS['stdev_fold'+str(PARAMS['fold'])])
fv_sp_patches = preproc.get_feature_patches(PARAMS, fv_sp, PARAMS['W'], PARAMS['W_shift'], featName)
if PARAMS['skewness_vector']:
if PARAMS['skewness_vector']=='Harm':
fv_sp_patches = cget_data_statistics(fv_sp_patches, axis=1)
fv_sp_patches = np.expand_dims(fv_sp_patches, axis=2)
elif PARAMS['skewness_vector']=='Perc':
fv_sp_patches = cget_data_statistics(fv_sp_patches, axis=0)
fv_sp_patches = np.expand_dims(fv_sp_patches, axis=1)
if balance_sp==0:
batchData_sp = fv_sp_patches
else:
batchData_sp = np.append(batchData_sp, fv_sp_patches, axis=0)
balance_sp += np.shape(fv_sp_patches)[0]
# print('Speech: ', batchSize, balance_sp, np.shape(batchData_sp))
while balance_mu<batchSize:
if not file_list_mu_temp:
file_list_mu_temp = file_list['music'].copy()
mu_fName = file_list_mu_temp.pop()
mu_fName_path = folder + '/music/' + mu_fName
if not os.path.exists(mu_fName_path):
# print(mu_fName_path, os.path.exists(mu_fName_path))
continue
fv_mu = preproc.get_featuregram(PARAMS, 'music', PARAMS['feature_opDir'], '', mu_fName_path, None, n_fft, n_mels, featName)
if PARAMS['frame_level_scaling']:
fv_mu = cscale_data(fv_mu, PARAMS['mean_fold'+str(PARAMS['fold'])], PARAMS['stdev_fold'+str(PARAMS['fold'])])
fv_mu_patches = preproc.get_feature_patches(PARAMS, fv_mu, PARAMS['W'], PARAMS['W_shift'], featName)
if PARAMS['skewness_vector']:
if PARAMS['skewness_vector']=='Harm':
fv_mu_patches = cget_data_statistics(fv_mu_patches, axis=1)
fv_mu_patches = np.expand_dims(fv_mu_patches, axis=2)
elif PARAMS['skewness_vector']=='Perc':
fv_mu_patches = cget_data_statistics(fv_mu_patches, axis=0)
fv_mu_patches = np.expand_dims(fv_mu_patches, axis=1)
if balance_mu==0:
batchData_mu = fv_mu_patches
else:
batchData_mu = np.append(batchData_mu, fv_mu_patches, axis=0)
balance_mu += np.shape(fv_mu_patches)[0]
# print('Music: ', batchSize, balance_mu, np.shape(batchData_mu))
batchData = batchData_mu[:batchSize, :] # music label=0
batchData = np.append(batchData, batchData_sp[:batchSize, :], axis=0) # speech label=1
balance_mu -= batchSize
balance_sp -= batchSize
batchData_mu = batchData_mu[batchSize:, :]
batchData_sp = batchData_sp[batchSize:, :]
batchLabel = [0]*batchSize # music
batchLabel.extend([1]*batchSize) # speech
batchLabel_smr = np.ones((3*batchSize,2))
batchLabel_smr[:batchSize, :] = np.repeat(np.array([1, 0], ndmin=2), batchSize, axis=0) # music
batchLabel_smr[batchSize:2*batchSize] = np.repeat(np.array([0, 1], ndmin=2), batchSize, axis=0) # speech
if len(PARAMS['classes'])==3:
while balance_spmu<batchSize:
if not file_list_spmu_temp:
file_list_spmu_temp = file_list['speech+music'].copy()
np.random.shuffle(file_list_spmu_temp)
spmu_info = file_list_spmu_temp.pop()
sp_fName = spmu_info['speech']
sp_fName_path = folder + '/speech/' + sp_fName
mu_fName = spmu_info['music']
mu_fName_path = folder + '/music/' + mu_fName
target_dB = spmu_info['SMR']
if (not os.path.exists(mu_fName_path)) or (not os.path.exists(sp_fName_path)):
continue
fv_spmu = preproc.get_featuregram(PARAMS, 'speech_music', PARAMS['feature_opDir'], sp_fName_path, mu_fName_path, target_dB, n_fft, n_mels, featName)
if PARAMS['frame_level_scaling']:
fv_spmu = cscale_data(fv_spmu, PARAMS['mean_fold'+str(PARAMS['fold'])], PARAMS['stdev_fold'+str(PARAMS['fold'])])
fv_spmu_patches = preproc.get_feature_patches(PARAMS, fv_spmu, PARAMS['W'], PARAMS['W_shift'], featName)
if PARAMS['skewness_vector']:
if PARAMS['skewness_vector']=='Harm':
fv_spmu_patches = cget_data_statistics(fv_spmu_patches, axis=1)
fv_spmu_patches = np.expand_dims(fv_spmu_patches, axis=2)
elif PARAMS['skewness_vector']=='Perc':
fv_spmu_patches = cget_data_statistics(fv_spmu_patches, axis=0)
fv_spmu_patches = np.expand_dims(fv_spmu_patches, axis=1)
if balance_spmu==0:
batchData_spmu = fv_spmu_patches
batchData_spmu_target_dB = np.array([target_dB]*np.shape(fv_spmu_patches)[0])
else:
batchData_spmu = np.append(batchData_spmu, fv_spmu_patches, axis=0)
batchData_spmu_target_dB = np.append(batchData_spmu_target_dB, np.array([target_dB]*np.shape(fv_spmu_patches)[0]))
balance_spmu += np.shape(fv_spmu_patches)[0]
# print('SpeechMusic: ', batchSize, balance_spmu, np.shape(batchData_spmu))
# speech_music label=2
batchData = np.append(batchData, batchData_spmu[:batchSize, :], axis=0)
balance_spmu -= batchSize
batchData_spmu = batchData_spmu[batchSize:, :]
batchLabel.extend([2]*batchSize) # speech+music
label_idx = 2*batchSize
for i in range(batchSize):
if batchData_spmu_target_dB[i]>=0:
batchLabel_smr[label_idx] = np.array([1/np.power(10,(batchData_spmu_target_dB[i]/10)), 1], ndmin=2)
else:
batchLabel_smr[label_idx] = np.array([1, np.power(10,(batchData_spmu_target_dB[i]/10))], ndmin=2)
label_idx += 1
batchData_spmu_target_dB = batchData_spmu_target_dB[batchSize:]
if 'Lemaire_et_al' in PARAMS['Model']:
batchData = np.transpose(batchData, axes=(0,2,1)) # TCN input shape=(batch_size, timesteps, ndim)
''' Adding Normal (Gaussian) noise for data augmentation '''
if PARAMS['data_augmentation_with_noise']:
scale = np.random.choice([5e-3, 1e-3, 5e-4, 1e-4])
noise = np.random.normal(loc=0.0, scale=scale, size=np.shape(batchData))
batchData = np.add(batchData, noise)
OHE_batchLabel = to_categorical(batchLabel, num_classes=len(PARAMS['classes']))
'''
Speech Nonspeech
'''
batchLabel_sp_nsp = np.copy(batchLabel)
batchLabel_sp_nsp[:batchSize] = 0
batchLabel_sp_nsp[batchSize:2*batchSize] = 1
batchLabel_sp_nsp[2*batchSize:] = 0
'''
Music Nonmusic
'''
batchLabel_mu_nmu = np.copy(batchLabel)
batchLabel_mu_nmu[:batchSize] = 1
batchLabel_mu_nmu[batchSize:2*batchSize] = 0
batchLabel_mu_nmu[2*batchSize:] = 0
batchLabel_MTL = {'R': batchLabel_smr, 'S': batchLabel_sp_nsp, 'M': batchLabel_mu_nmu, '3C': OHE_batchLabel}
batch_count += 1
# print('Batch ', batch_count, ' shape=', np.shape(batchData), np.shape(OHE_batchLabel))
if ('MTL' in PARAMS['Model']) or ('Cascaded_MTL' in PARAMS['Model']):
yield batchData, batchLabel_MTL
else:
yield batchData, OHE_batchLabel