def PackageFeatAndAli(all_package, input_lock, package_end, scp_file, ali_file, nstreams, skip_frame = 1, max_input_seq_length = 1500, criterion = 'ce'):
logging.info('------start PackageFeatAndAli------')
start_package = time.time()
#all_package = []
# first read ali
alignment_dict = read_alignment(ali_file)
scp_list = []
ali_list = []
# second read feature scp file and package feature and ali
for line in open(scp_file, 'r'):
utt_id, utt_mat = read_next_utt(line)
logging.debug(utt_id + ' read ok.')
# overlength
if int(len(utt_mat)/skip_frame) + 1 > max_input_seq_length:
logging.info(utt_id + ' length '+ str(int(len(utt_mat)/skip_frame)+1) + ' > ' + str(max_input_seq_length))
continue
try:
ali_utt = alignment_dict[utt_id]
except KeyError:
logging.info('no '+ utt_id + ' align')
continue
if 'ce' in criterion:
if len(utt_mat) != len(ali_utt):
# delete one frame ali in utt_mat
if len(ali_utt) - len(utt_mat) == 1:
#logging.warn(utt_id + ' feat length + 1 = ali length , and delete one ali label')
ali_utt = numpy.delete(ali_utt, -1)
else:
logging.info(utt_id + ' feat and ali isn\'t equal length')
continue
elif 'ctc' in criterion :
if len(utt_mat) < len(ali_utt) * 2 - 1:
logging.info(utt_id + ' feat < ali * 2 - 1 :%d < %d * 2 - 1' % (len(utt_mat), len(ali_utt)))
continue
scp_list.append(line)
ali_list.append(ali_utt)
if len(scp_list) == nstreams:
input_lock.acquire()
all_package.append([scp_list, ali_list])
input_lock.release()
scp_list = []
ali_list = []
if len(scp_list) != 0:
while len(scp_list) < nstreams:
scp_list.append(scp_list[0])
ali_list.append(ali_list[0])
input_lock.acquire()
all_package.append([scp_list, ali_list])
input_lock.release()
input_lock.acquire()
package_end.append(True)
input_lock.release()
end_package = time.time()
logging.info('------PackageFeatAndAli end. Package time is : %f s, batch number : %d' % (end_package - start_package, len(all_package)))
return True
def PackageFeatAndAliAndLat(all_package, input_lock, package_end, feat_scp_file, ali_file, lat_scp_file, nstreams,
skip_frame = 1, max_input_seq_length = 1500, criterion = 'mmi'):
logging.info('------start PackageFeatAndAliAndLat------')
start_package = time.time()
# first read ali
alignment_dict = read_alignment(ali_file)
lat_dict = ReadScp(lat_scp_file)
feat_list = []
ali_list = []
lat_list = []
with open(feat_scp_file, 'r') as feat_fp:
for line in feat_fp:
utt_id, utt_mat = read_next_utt(line)
logging.debug(utt_id + ' read ok.')
if int(len(utt_mat)/skip_frame) + 1 > max_input_seq_length:
logging.info(utt_id + ' length '+ str(int(len(utt_mat)/skip_frame)+1) + ' > ' + str(max_input_seq_length))
continue
try:
ali_utt = alignment_dict[utt_id]
except KeyError:
logging.info('no '+ utt_id + ' align')
continue
try:
lat_scp_line = lat_dict[utt_id]
except KeyError:
logging.info('no '+ utt_id + ' lattice')
continue
# should check length is equal.
# but because of time question , now it's not check length.
feat_list.append(line)
ali_list.append(ali_utt)
lat_list.append(lat_scp_line)
if len(feat_list) == nstreams:
input_lock.acquire()
all_package.append([feat_list, ali_list, lat_list])
input_lock.release()
feat_list = []
ali_list = []
lat_list = []
if len(feat_list) != 0:
while len(feat_list) < nstreams:
feat_list.append(feat_list[0])
ali_list.append(ali_list[0])
lat_list.append(lat_list[0])
input_lock.acquire()
all_package.append([feat_list, ali_list, lat_list])
input_lock.release()
input_lock.acquire()
package_end.append(True)
input_lock.release()
end_package = time.time()
logging.info('------PackageFeatAndAliAndLat end. Package time is : %f s, batch number : %d' % (end_package - start_package, len(all_package)))
def LoadOnePackage(self):
while True:
self.input_lock.acquire()
if self.read_offset.value >= len(self.package_feat_ali):
if self.package_end[-1] is True:
self.input_lock.release()
return None, None, None, None, None
else:
self.input_lock.release()
time.sleep(0.05)
continue
else:
package = self.package_feat_ali[self.read_offset.value]
self.read_offset.value += 1
self.input_lock.release()
break
feat_scp = package[0]
label = package[1]
lat_scp = None
lat_list = None
if len(package) == 3:
lat_scp = package[2]
# indexs_info_list, pdf_values_list, lmweight_values_list, amweight_values_list, statesinfo_list, statenum_list, time_list
lat_list = PackageLattice(lat_scp, map_pdf_phone = self.ali_to_pdf_phone)
max_frame_num = 0
length = []
feat_mat = []
for feat_line in feat_scp:
utt_id, utt_mat = read_next_utt(feat_line)
# do feature transform
if self.feature_transform != None:
utt_mat = self.feature_transform.Propagate(utt_mat)
feat_mat.append(
skip_frame(utt_mat ,self.skip_frame, self.skip_offset))
length.append(len(feat_mat[-1]))
if max_frame_num < length[-1]:
max_frame_num = length[-1]
if self.do_skip_lab and self.skip_frame > 1:
process_lab = []
for lab in label:
# if self.skip_frame > 1:
process_lab.append(lab[self.skip_offset : len(lab) : self.skip_frame])
label = process_lab
# check lattice and label and feat
if lat_list is not None:
i = 0
lat_times = lat_list[-1]
while i < len(lat_times):
assert lat_times[i] == length[i]
i += 1
return feat_mat, label, length, max_frame_num, lat_list
def LoadOnePackage(self):
while True:
self.input_lock.acquire()
if self.read_offset >= len(self.package_feat_ali):
if self.package_end[-1] is True:
self.input_lock.release()
return None, None, None, None
else:
self.input_lock.release()
time.sleep(1.0)
continue
else:
self.input_lock.release()
break
package = self.package_feat_ali[self.read_offset]
self.read_offset += 1
feat_scp = package[0]
label = package[1]
max_frame_num = 0
length = []
feat_mat = []
for feat_line in feat_scp:
utt_id, utt_mat = read_next_utt(feat_line)
# do feature transform
if self.feature_transform != None:
utt_mat = self.feature_transform.Propagate(utt_mat)
feat_mat.append(
skip_frame(utt_mat, self.skip_frame, self.skip_offset))
length.append(len(feat_mat[-1]))
if max_frame_num < length[-1]:
max_frame_num = length[-1]
if self.do_skip_lab and self.skip_frame > 1:
process_lab = []
for lab in label:
# if self.skip_frame > 1:
process_lab.append(
lab[self.skip_offset:len(lab):self.skip_frame])
label = process_lab
return feat_mat, label, length, max_frame_num