def _read_output(self):
for i, fname in enumerate(self._output_file_list): \
fname = fname.split('num')[-1].split('.raw')[0]
if self._dataindx == fname:
# print(self._output_file_list[i])
dataname = os.path.dirname(self._output_file_list[i]) + \
'/' + os.path.basename(self._output_file_list[i]).split('.')[0] + '.bin'
if os.path.exists(dataname):
feat_spec = np.load(dataname.replace('.bin',
'.npy')).item()
feat_shape = feat_spec['shape']
feat_max = feat_spec['max']
feat = utils.read_raw(dataname).reshape(
feat_shape) * np.float32(feat_max)
else:
data = utils.read_raw(self._output_file_list[i])
if config.parallel:
feat = self.lpsd_dist_p(
data, self._dist_num, is_patch=False
) # (The number of samples, config.freq_size, 1, 1)
else:
feat = self.lpsd_dist(
data, self._dist_num, is_patch=False
) # (The number of samples, config.freq_size, 1, 1)
utils.write_bin(feat, np.max(np.abs(feat)), dataname)
feat_spec = {
'shape': feat.shape,
'max': np.max(np.abs(feat))
}
np.save(dataname.replace('.bin', ''), feat_spec)
# plt.subplot(212)
# S = librosa.amplitude_to_db(
# librosa.stft(data, hop_length=config.win_step,
# win_length=config.win_size, n_fft=config.nfft), ref=np.max)
# ld.specshow(S, y_axis='linear', hop_length=config.win_step, sr=config.fs)
#
# plt.show()
# data, _ = librosa.load(self._output_file_list[i], config.fs)
break
# print('output')
# if self._is_shuffle:
#
# feat = np.reshape(feat, (-1, self._batch_size, feat.shape[1], feat.shape[2], feat.shape[3]))
# feat = np.reshape(feat[self._perm_indx, :], (-1, config.freq_size, 1, 1))
''''''
if self._is_shuffle:
feat = feat[self._perm_indx, :]
return np.squeeze(feat)
def format(self, raw):
'''
Raw is the output of the psycopg2 query, a list of lists of lists etc...
'''
data = Data()
data.key = [row[0] for row in raw]
data.data_id = [row[1] for row in raw]
data.filename = [row[3] for row in raw]
#The follwing writes each binary to a file
for row in raw:
write_bin(self.write_bin_dir + row[3], row[2])
return data
def face_register(self, cvImg, faceId):
save_path = self.faceDataDir + str(faceId)
if not os.path.exists(save_path):
os.mkdir(save_path)
retDict = {}
featFn = save_path + '/feat.bin'
feat_dict = self.face_feature(cvImg)
if feat_dict['flag'] != 'failure':
write_bin(feat_dict['feature'], featFn)
retDict['flag'] = 'successful'
else:
retDict['flag'] = 'feature'
imgFn = save_path + '/face.jpg'
cv2.imwrite(imgFn, cvImg)
return retDict
def prepare_args(input_no, poc, poc_fmt):
global TmpFolder
# prepare the arguments
arg_num = len(poc_fmt)
arg_list = []
for arg_no in range(arg_num):
if poc_fmt[arg_no][0] == 'bfile': # write the list into binary file
content = np.asarray(poc[poc_fmt[arg_no][1]: poc_fmt[arg_no][1]+poc_fmt[arg_no][2]]).astype(np.int)
tmp_filepath = os.path.join(TmpFolder, 'tmp_%d' % input_no)
utils.write_bin(tmp_filepath, content)
arg_list.append(tmp_filepath)
elif poc_fmt[arg_no][0] == 'int':
arg_list.append(int(poc[poc_fmt[arg_no][1]]))
elif poc_fmt[arg_no][0] == 'float':
arg_list.append(float(poc[poc_fmt[arg_no][1]]))
elif poc_fmt[arg_no][0] == 'str': # concatenate all the chars together
arg_list.append(''.join(poc[poc_fmt[arg_no][1]: poc_fmt[arg_no][1]+poc_fmt[arg_no][2]]))
else:
raise Exception("ERROR: Unknown poc_fmt -> %s" % poc_fmt[arg_no][0])
return arg_list
def _read_input(self, input_file_dir):
# print("num_file: %.2d" % self._num_file)
# self._dataindx = input_file_dir.split("num")[1].split('.')[0]
# dataname = os.path.dirname(input_file_dir) + '/' + os.path.basename(input_file_dir).split('.')[0] + '.npy'
if self._is_training:
# print("num_file: %.3d" % self._num_file)
self._dataindx = input_file_dir.split("num")[1].split('.')[0]
dataname = os.path.dirname(
input_file_dir) + '/' + os.path.basename(input_file_dir).split(
'.')[0] + '.bin'
if os.path.exists(dataname):
# print("num_file: %.2d load..." % self._num_file)
feat_spec = np.load(dataname.replace('.bin', '.npy')).item()
feat_size = feat_spec['feat_size']
feat_shape = feat_spec['feat_shape']
phase_shape = feat_spec['phase_shape']
feat_max = feat_spec['max']
feat_phase = utils.read_raw(dataname) * np.float32(feat_max)
feat = feat_phase[0:feat_size].reshape(feat_shape)
phase = feat_phase[feat_size:].reshape(phase_shape)
# print("num_file: %.2d load... done" % self._num_file)
else:
print("num_file: %.2d feature extraction..." % self._num_file)
# data, _ = librosa.load(input_file_dir, config.fs)
data = utils.read_raw(input_file_dir)
# plt.subplot(211)
#
# S = librosa.amplitude_to_db(
# librosa.stft(data, hop_length=config.win_step,
# win_length=config.win_size, n_fft=config.nfft), ref=np.max)
# ld.specshow(S, y_axis='linear', hop_length=config.win_step, sr=config.fs)
if config.parallel:
feat, phase = self.lpsd_dist_p(data,
self._dist_num,
is_patch=True)
else:
feat, phase = self.lpsd_dist(data,
self._dist_num,
is_patch=True)
feat_size = feat.size
feat_phase = np.concatenate(
(feat.reshape(-1), phase.reshape(-1)))
utils.write_bin(feat_phase, np.max(np.abs(feat_phase)),
dataname)
feat_spec = {
'feat_size': feat_size,
'phase_shape': phase.shape,
'feat_shape': feat.shape,
'max': np.max(np.abs(feat_phase))
}
np.save(dataname.replace('.bin', ''), feat_spec)
print("num_file: %.2d feature extraction... done." %
self._num_file)
else:
data = np.load(input_file_dir)
# data, _ = librosa.load(input_file_dir, config.fs)
if config.parallel:
feat, phase = self.lpsd_dist_p(data, self._dist_num)
else:
feat, phase = self.lpsd_dist(data, self._dist_num)
# feat shape: (num samples, config.time_width, config.freq_size, 1)
# if self._is_shuffle:
# feat = np.reshape(feat, (-1, self._batch_size, feat.shape[1], feat.shape[2], feat.shape[3]))
# self._perm_indx = perm_indx = np.random.permutation(feat.shape[0])
# feat = np.reshape(feat[perm_indx, :], (-1, config.time_width, config.freq_size, 1))
''''''
if self._is_shuffle:
self._perm_indx = perm_indx = np.random.permutation(feat.shape[0])
feat = feat[perm_indx, :]
self.num_samples = np.shape(feat)[0]
if self._is_val:
phase = np.zeros(feat.shape)
return feat, phase
fs=config.fs)
# plt.plot(recon_speech)
# plt.show()
# lab = np.reshape(np.asarray(lab), [-1, 1])
test_dr.reader_initialize()
break
return recon_speech
if __name__ == '__main__':
input_dir = os.path.abspath('./data/train/noisy')
logs_dir = os.path.abspath('./logs' + '/logs_' +
time.strftime("%Y-%m-%d-%H-%M-%S"))
model_dir = os.path.abspath('./enhanced_wav/enhance_model')
save_dir = os.path.abspath('./enhanced_wav')
gs.freeze_graph(logs_dir, model_dir,
'model_1/pred,model_1/labels,model_1/cost')
input_file_list = sorted(glob.glob(input_dir + '/*.raw'))
graph_name = sorted(glob.glob(save_dir + '/*.pb'))[-1]
norm_path = os.path.abspath('./data/train/norm')
for wav_dir in input_file_list:
recon_speech = speech_enhance(wav_dir, graph_name)
file_dir = save_dir + '/' + os.path.basename(wav_dir).replace(
'noisy', 'enhanced')
utils.write_bin(recon_speech, np.max(np.abs(recon_speech)), file_dir)