def __getitem__(self, index):
if not self.io_status:
return 0., 0.
chunk = [
int(self.data_frame[index][2]),
int(self.data_frame[index][3])
]
egs = kaldi_io.read_mat(self.data_frame[index][1], chunk=chunk)
target = self.data_frame[index][4]
alignment = kaldi_io.read_ali_ark(self.data_frame[index][5],
chunk=chunk) #Zheng Li 2020-10
chunk_size = int(chunk[1] - chunk[0] + 1)
alignment = np.ascontiguousarray(alignment)
alignment = np.reshape(alignment, (chunk_size, 1))
# Note, egs read from kaldi_io is read-only and
# use egs = np.require(egs, requirements=['O', 'W']) to make it writeable.
# It avoids the problem "ValueError: assignment destination is read-only".
# Note that, do not use inputs.flags.writeable = True when the version of numpy >= 1.17.
egs = np.require(egs, requirements=['O', 'W'])
if self.aug is not None:
return self.aug(egs.T), target, alignment
else:
return egs.T, target, alignment
def __getitem__(self, index):
if not self.io_status:
return 0., 0.
chunk = [
int(self.data_frame[index][2]),
int(self.data_frame[index][3])
]
egs = kaldi_io.read_mat(self.data_frame[index][1], chunk=chunk)
if self.num_target_types == 1:
target = self.data_frame[index][4]
else:
target = self.data_frame[index][4:]
# Note, egs read from kaldi_io is read-only and
# use egs = np.require(egs, requirements=['O', 'W']) to make it writeable.
# It avoids the problem "ValueError: assignment destination is read-only".
# Note that, do not use inputs.flags.writeable = True when the version of numpy >= 1.17.
egs = np.require(egs, requirements=['O', 'W'])
if self.aug is not None:
return self.aug(egs.T), target
else:
return egs.T, target
def __getitem__(self, index):
if not self.io_status:
return 0., 0.
# Decode string from bytes after using astype(np.string_).
egs_path = str(self.ark_path[index][0], encoding='utf-8')
if self.chunk_position is not None:
chunk = [
self.chunk_position[index][0], self.chunk_position[index][1]
]
else:
chunk = None
if self.egs_type is "chunk":
egs = kaldi_io.read_mat(egs_path, chunk=chunk)
else:
egs = kaldi_io.read_vec_flt(egs_path)
target = self.label[index][0]
# Note, egs which is read from kaldi_io is read-only and
# use egs = np.require(egs, requirements=['O', 'W']) to make it writeable.
# It avoids the problem "ValueError: assignment destination is read-only".
# Note that, do not use inputs.flags.writeable = True when the version of numpy >= 1.17.
egs = np.require(egs, requirements=['O', 'W'])
if self.aug is not None:
return self.aug(egs.T), target
else:
return egs.T, target
def get_base_attribute_(self):
## Base attribute
# Total utts
self.num_utts = len(self.utt2spk) if "utt2spk" in self.loaded_attr else None
# Total spks
self.num_spks = len(self.spk2utt) if "spk2utt" in self.loaded_attr else None
# Total frames
if "utt2num_frames" in self.loaded_attr:
self.num_frames = 0
for utt, num_frames in self.utt2num_frames.items():
self.num_frames += num_frames
else:
self.num_frames = None
# Feature dim
self.feat_dim = kaldi_io.read_mat(
self.feats_scp[list(self.feats_scp.keys())[0]]).shape[1] if "feats_scp" in self.loaded_attr else None
# Select device
model = utils.select_model_device(model, args.use_gpu, gpu_id=args.gpu_id)
model.eval()
with kaldi_io.open_or_fd(args.feats_rspecifier, "rb") as r, \
kaldi_io.open_or_fd(args.vectors_wspecifier, 'wb') as w:
for line in r:
# (key, rxfile, chunk_start, chunk_end) = line.decode().split(' ')
# chunk=[chunk_start, chunk_end]
# print("Process utterance for key {0}".format(key))
# feats = kaldi_io.read_mat(rxfile, chunk=chunk)
(key, rxfile) = line.decode().split(' ')
print("Process utterance for key {0}".format(key))
feats = kaldi_io.read_mat(rxfile)
embedding = model.extract_embedding(feats)
kaldi_io.write_vec_flt(w, embedding.numpy(), key=key)
# while(True):
# key = kaldi_io.read_key(r)
# if not key:
# break
# print("Process utterance for key {0}".format(key))
# feats = kaldi_io.read_mat(r)
# embedding = model.extract_embedding(feats)
# kaldi_io.write_vec_flt(w, embedding.numpy(), key=key)
except BaseException as e:
if not isinstance(e, KeyboardInterrupt):
traceback.print_exc()
)
model = utils.create_model_from_py(model_blueprint, model_creation)
model.load_state_dict(torch.load(args.model_path, map_location='cpu'),
strict=False)
# Select device
model = utils.select_model_device(model, args.use_gpu, gpu_id=args.gpu_id)
model.eval()
with kaldi_io.open_or_fd(args.feats_rspecifier, "rb") as r, \
kaldi_io.open_or_fd(args.vectors_wspecifier, 'wb') as w:
while (True):
key = kaldi_io.read_key(r)
if not key:
break
print("Process utterance for key {0}".format(key))
feats = kaldi_io.read_mat(r)
embedding = model.extract_embedding(feats)
kaldi_io.write_vec_flt(w, embedding.numpy(), key=key)
except BaseException as e:
if not isinstance(e, KeyboardInterrupt):
traceback.print_exc()
sys.exit(1)