def file_reader(record_set_dir, batch_size=-1, is_shuffle=True):
""" read wav files in given directies, one file per time
Args:
record_set_dir: directory or list of directories where recordings exist
Returns:
samples generator, [samples, label_all]
"""
if isinstance(record_set_dir, list):
dirs = record_set_dir
else:
dirs = [record_set_dir]
#
fpath_all = []
for sub_set_dir in dirs:
fpath_all_sub = get_fpath(sub_set_dir, '.wav', is_absolute=True)
fpath_all.extend(fpath_all_sub)
if is_shuffle:
np.random.shuffle(fpath_all)
# print('#file',len(fpath_all))
# raise Exception()
if len(fpath_all) < 1:
raise Exception('empty folder:{}'.format(record_set_dir))
frame_len = 320
shift_len = 160
n_azi = 37
if batch_size > 1:
x_all = np.zeros((0, frame_len, 2, 1))
y_all = np.zeros((0, n_azi))
for fpath in fpath_all:
record, fs = wav_tools.read_wav(fpath)
x_file_all = wav_tools.frame_data(record, frame_len, shift_len)
x_file_all = np.expand_dims(x_file_all, axis=-1)
# onehot azi label
n_sample_file = x_file_all.shape[0]
fname = os.path.basename(fpath)
azi = np.int16(fname.split('_')[0])
y_file_all = np.zeros((n_sample_file, n_azi))
y_file_all[:, azi] = 1
if batch_size > 0:
x_all = np.concatenate((x_all, x_file_all), axis=0)
y_all = np.concatenate((y_all, y_file_all), axis=0)
while x_all.shape[0] > batch_size:
x_batch = copy.deepcopy(x_all[:batch_size])
y_batch = copy.deepcopy(y_all[:batch_size])
x_all = x_all[batch_size:]
y_all = y_all[batch_size:]
yield [x_batch, y_batch]
else:
yield [x_file_all, y_file_all]
def cal_fea(record_dir, fea_dir):
"""calculate GCC-PHAT features
Args:
record_dir: wave dataset directory
"""
if not os.path.exists(record_dir):
os.makedirs(record_dir)
wav_fpath_all = get_fpath(dir_path=record_dir,
suffix='.wav',
pattern='reverb')
pb = ProcessBar(max_value=len(wav_fpath_all),
title=f'GCC_PHAT {record_dir}')
pool = Pool(24)
for wav_fpath in wav_fpath_all:
fea_fpath = os.path.join(fea_dir, '{}.npy'.format(wav_fpath[:-4]))
if os.path.exists(fea_fpath):
# warnings.warn(f'{fea_fpath} exists!')
continue
data, fs = wav_tools.read_wav(os.path.join(record_dir, wav_fpath))
frame_all = wav_tools.frame_data(data, frame_len=320, shift_len=160)
n_frame = frame_all.shape[0]
fea_frame_all = pool.map(gcc_phat_parallel_f,
[frame_all[i] for i in range(n_frame)])
fea_frame_all = np.asarray(fea_frame_all)
dir_tmp = os.path.dirname(fea_fpath)
if not os.path.exists(dir_tmp):
os.makedirs(dir_tmp)
np.save(fea_fpath, fea_frame_all)
pb.update()
def file_reader(reverb_set_dir,
batch_size=128,
is_shuffle=True,
frame_len=320,
shift_len=160,
n_azi=37):
""" read wav files in given directies, one file per time
Args:
record_set_dir: directory or list of directories where recordings exist
batch_size:
is_shuffle:
Returns:
samples generator, [samples, label_all]
"""
if isinstance(reverb_set_dir, list):
dir_all = reverb_set_dir
else:
dir_all = [reverb_set_dir]
#
fpath_reverb_all = []
for dir_fpath in dir_all:
fpath_all_tmp = get_fpath(dir_fpath, '.npy', is_absolute=True)
fpath_reverb_all.extend(fpath_all_tmp)
if is_shuffle:
np.random.shuffle(fpath_reverb_all)
for fpath_reverb in fpath_reverb_all:
x_d_batch, x_r_batch, y_loc_batch, is_anechoic = np.load(
fpath_reverb, allow_pickle=True)
# if x_d.shape[0] == batch_size and x_r.shape[0] == batch_size and y_loc.shape[0] == batch_size:
yield x_r_batch, y_loc_batch
def file_reader(dataset_dir, norm_coef_fpath, batch_size=-1, is_shuffle=True):
"""Read spectrum files under given directory
Args:
dataset_dir: string or list of strings
norm_coef_fpath: file path of normalization coefficients
batch_size: if not specified, return the data of a file per time
is_shuffle:
Returns:
sample,label_onehot
"""
n_azi = 37 # number of sound position
fea_len = 37 # size of feature
if isinstance(dataset_dir, list):
dir_all = dataset_dir
else:
dir_all = [dataset_dir]
mean, std = np.load(norm_coef_fpath) #
fpath_all = []
for dir_tmp in dir_all:
fpath_all_tmp = get_fpath(dir_tmp, suffix='.npy', is_absolute=True)
fpath_all.extend(fpath_all_tmp)
if is_shuffle:
np.random.shuffle(fpath_all) # randomize files order
if len(fpath_all) < 1:
raise Exception('folder is empty: {}'.format(dataset_dir))
if batch_size > 0:
x_all = np.zeros((0, fea_len))
y_all = np.zeros((0, n_azi))
for fpath in fpath_all:
fea_file_all = np.load(fpath)
x_file_all = np.divide(fea_file_all - mean, std)
n_sample_tmp = x_file_all.shape[0]
fname, _ = os.path.basename(fpath).split('.')
azi, _ = map(int, fname.split('_'))
y_file_all = np.zeros((n_sample_tmp, n_azi))
y_file_all[:, azi] = 1
if batch_size > 0: #
x_all = np.concatenate((x_all, x_file_all), axis=0)
y_all = np.concatenate((y_all, y_file_all), axis=0)
while (x_all.shape[0] > batch_size):
x_batch = copy.deepcopy(x_all[:batch_size])
y_batch = copy.deepcopy(y_all[:batch_size])
x_all = x_all[batch_size:]
y_all = y_all[batch_size:]
yield [x_batch, y_batch]
else:
yield [x_file_all, y_file_all]
def file_reader(data_dir):
"""Read spectrum files under given directory
Args:
data_dir:
Returns:
sample generator, [samples,label_onehots]
"""
if isinstance(data_dir, list):
dir_list = data_dir
else:
dir_list = [data_dir]
for sub_set_dir in dir_list:
fea_fpath_list = get_fpath(sub_set_dir, '.npy')
for fea_fpath in fea_fpath_list:
fea_fpath_abs = os.path.join(sub_set_dir, fea_fpath)
fea = np.load(fea_fpath_abs)
yield fea
def load_src_fpath(record_dir):
src_fpath_fpath_all = get_fpath(record_dir,
suffix='.txt',
is_absolute=True)
fpath_all = {}
for src_fpath_fpath in src_fpath_fpath_all:
*_, room, mic_pos, fname = src_fpath_fpath.split('/')
if fname != 'src_fpath.txt':
raise Exception
if room not in fpath_all.keys():
fpath_all[room] = {}
if mic_pos not in fpath_all[room].keys():
fpath_all[room][mic_pos] = {}
with open(src_fpath_fpath, 'r') as src_fpath_file:
line_all = src_fpath_file.readlines()
for line in line_all[1:]:
condition, fpath_tar, fpath_inter = line.split()
# print(condition, fpath_tar, fpath_inter)
fpath_all[room][mic_pos][condition] = [fpath_tar, fpath_inter]
return fpath_all
azi_tar_all = np.arange(8, 29)
n_inter_all = [0, 1, 2, 3]
n_test = 3
n_room = len(room_all)
n_mic_pos = len(mic_pos_test_all)
n_azi_tar = azi_tar_all.shape[0]
n_n_inter = len(n_inter_all)
record_set_dir = '../Data/Records/test'
fea_set_dir = '../Data/Features/test'
# wave file path of TIMIT test set
TIMIT_dir = '/home/st/Work_Space/Data/TIMIT/44100Hz/TIMIT/TEST'
src_fpath_all = [
os.path.join(TIMIT_dir, item) for item in get_fpath(TIMIT_dir, '.wav')
]
def get_wav_fpath(n):
return np.random.choice(src_fpath_all, size=n, replace=False)
def syn_record(src, room, mic_pos, azi, filter_gpu):
brir_fpath = '../Data/BRIRs/test/{}/{}/{}.mat'.format(room, mic_pos, azi)
brir = sio.loadmat(brir_fpath)['data']
return filter_gpu.brir_filter(src, brir)
def gen_test_sample(room, mic_pos, azi_tar, n_inter, test_i, filter_gpu,
front_end):
def wav2npy(reverb_set_dir, npy_dir, is_anechoic):
""" read wav files in given directies, one file per time
Args:
record_set_dir: directory or list of directories where recordings exist
batch_size:
is_shuffle:
Returns:
samples generator, [samples, label_all]
"""
frame_len = 320
shift_len = 160
n_azi = 37
batch_size = 128
os.makedirs(npy_dir, exist_ok=True)
#
fpath_reverb_all = get_fpath(reverb_set_dir, '.wav', is_absolute=True)
if len(fpath_reverb_all) < 1:
raise Exception('empty folder:{}'.format(reverb_set_dir))
pb = ProcessBar(len(fpath_reverb_all))
batch_count = 0
x_r = np.zeros((0, frame_len, 2, 1))
x_d = np.zeros((0, frame_len, 2, 1))
y_loc = np.zeros((0, n_azi))
for fpath_reverb in fpath_reverb_all:
pb.update()
# reverb signal
record, fs = wav_tools.read_wav(fpath_reverb)
x_r_file = np.expand_dims(wav_tools.frame_data(record, frame_len,
shift_len),
axis=-1)
# direct signal
fpath_direct = fpath_reverb.replace('reverb', 'direct')
direct, fs = wav_tools.read_wav(fpath_direct)
x_d_file = np.expand_dims(wav_tools.frame_data(direct, frame_len,
shift_len),
axis=-1)
# onehot azi label
n_sample_file = x_d_file.shape[0]
if x_r_file.shape[0] != n_sample_file:
raise Exception('sample number do not consist')
fname = os.path.basename(fpath_reverb)
azi = np.int16(fname.split('_')[0])
y_loc_file = np.zeros((n_sample_file, n_azi))
y_loc_file[:, azi] = 1
x_r = np.concatenate((x_r, x_r_file), axis=0)
x_d = np.concatenate((x_d, x_d_file), axis=0)
y_loc = np.concatenate((y_loc, y_loc_file), axis=0)
while x_d.shape[0] > batch_size:
x_r_batch = x_r[:batch_size]
x_d_batch = x_d[:batch_size]
y_loc_batch = y_loc[:batch_size]
npy_fpath = os.path.join(npy_dir, '{}.npy'.format(batch_count))
np.save(npy_fpath,
[x_d_batch, x_r_batch, y_loc_batch, is_anechoic])
batch_count = batch_count + 1
x_r = x_r[batch_size:]
x_d = x_d[batch_size:]
y_loc = y_loc[batch_size:]
def file_reader(fea_dir,
band_tar=None,
azi_tar=None,
is_screen=False,
record_dir=None,
is_plot=False,
fig_name=None,
is_pb=False):
#
theta_vad = 40
theta_corr_coef = 0.3
theta_itd = 44.0 / 44.1
if is_screen:
src_fpath_all = load_src_fpath(record_dir)
fea_fpath_all = get_fpath(fea_dir, suffix='.npz', is_absolute=True)
pb = ProcessBar(len(fea_fpath_all))
for fea_fpath in fea_fpath_all:
if is_pb:
pb.update()
*_, room, mic_pos, fname = fea_fpath[:-4].split('/')
azi, wav_i, snr = [np.int16(item) for item in fname.split('_')]
if (azi_tar is not None) and (azi != azi_tar):
continue
fea_file = np.load(fea_fpath)
cue_frame_all = fea_file['cue_frame_all']
ccf_frame_all = fea_file['ccf_frame_all']
snr_frame_all = fea_file['snr_frame_all']
if not is_screen:
if band_tar is None:
yield np.transpose(cue_frame_all, axes=(1, 0, 2))
else:
yield cue_frame_all[band_tar]
else:
n_frame = cue_frame_all.shape[1]
flag_frame_all_band_all = []
# feature selection
# 1. vad, on one channel(L)
src_fpath_tar = \
src_fpath_all[room][mic_pos][f'{azi}_{wav_i}_{snr}'][0]
src_fpath_tar = src_fpath_tar.replace('Data/TIMIT',
'Data/TIMIT_wav')
src_tar, fs = wav_tools.read_wav(src_fpath_tar)
tar_fpath = ''.join((f'{record_dir}/{room}/{mic_pos}/',
f'{azi}_{wav_i}_{snr}_tar.wav'))
tar, fs = wav_tools.read_wav(tar_fpath)
if tar.shape[0] != src_tar.shape[0]:
raise Exception()
# time delay between source and recording, about 70 samples
delay = 190
src_tar = np.concatenate((src_tar[delay:], np.zeros(delay)))
vad_flag_frame_all = wav_tools.vad(x=src_tar,
frame_len=frame_len,
shift_len=shift_len,
theta=theta_vad,
is_plot=False)
vad_flag_frame_all = np.squeeze(vad_flag_frame_all[:n_frame])
for band_i in range(n_band):
if band_tar is not None:
if band_i != band_tar:
continue
# 2. SNR in each frequency band
snr_flag_frame_all = snr_frame_all[band_i] > 0.0
# 3. correlation coefficients
ccf_flag_frame_all = np.greater(
np.max(ccf_frame_all[band_i], axis=1), theta_corr_coef)
# 4. ITDs range
itd_flag_frame_all = np.less(
np.abs(cue_frame_all[band_i, :, 0]), theta_itd) # itd ms
# combine all criteras
flag_frame_all = np.logical_and.reduce(
(vad_flag_frame_all, snr_flag_frame_all,
ccf_flag_frame_all, itd_flag_frame_all))
flag_frame_all_band_all.append(flag_frame_all)
# plot waveform and corresponding criteria result
if is_plot:
tar_fpath = os.path.join('{}_{}.wav'.format(azi, wav_i))
tar, fs = wav_tools.read_wav(tar_fpath)
inter_fpath = '{}_{}_{}.wav'.format(azi, wav_i, snr)
inter, fs = wav_tools.read_wav(inter_fpath)
front_end = Auditory_model(fs=fs,
cf_low=freq_low,
cf_high=freq_high,
n_band=n_band,
is_middle_ear=True,
ihc_type='Roman')
t_frame = np.arange(n_frame) * shift_len + int(frame_len / 2)
fig = plt.figure(figsize=(8, 4), tight_layout=True)
ax1 = plt.subplot(221)
ax1.plot(np.sum(front_end.filter(inter)[band_i], axis=1))
ax1.plot(np.sum(front_end.filter(tar)[band_i], axis=1))
ax_twin = ax1.twinx()
ax_twin.plot(t_frame, flag_frame_all, color='red')
ax2 = plt.subplot(223)
ax2.plot(t_frame, vad_flag_frame_all + 0.09, label='vad')
ax2.plot(t_frame, snr_flag_frame_all + 0.06, label='snr')
ax2.plot(t_frame, ccf_flag_frame_all + 0.03, label='ccf')
ax2.plot(t_frame, itd_flag_frame_all, label='itd')
ax2.legend()
ax3 = plt.subplot(122)
plot_cue_sample(cue_frame_all[band_i], ax3)
plot_cue_sample(cue_frame_all[band_i, flag_frame_all, :], ax3)
plot_tools.savefig(fig, fig_name=fig_name, fig_dir='./')
return
if band_i is None:
flag_frame_all = np.logical_and.reduce(flag_frame_all_band_all)
yield np.transpose(cue_frame_all[:, flag_frame_all, :],
axes=(1, 0, 2))
else:
flag_frame_all = flag_frame_all_band_all[0]
yield cue_frame_all[band_tar, flag_frame_all, :]