def LoadBatch(self):
while True:
if 'cnn' in self.criterion:
if 'whole' in self.criterion or 'ctc' in self.criterion:
feat,label,length,lattice = self.CnnLoadNextNstreams()
else:
feat,label,length,lattice = self.CnnSliceLoadNextNstreams()
elif 'tdnn' in self.criterion:
feat,label,length,lattice = self.TdnnLoadNextNstreams()
elif 'chain' in self.criterion:
feat,label,length,lattice = self.ChainLoadNextNstreams()
else:
if 'whole' in self.criterion or 'ctc' in self.criterion:
feat,label,length,lattice = self.WholeLoadNextNstreams()
else:
feat,label,length,lattice = self.SliceLoadNextNstreams()
if label is not None:
if 'ctc' in self.criterion:
label = sparse_tuple_from(label)
self.input_queue.put((feat,label,length,lattice))
if feat is not None:
print(numpy.shape(feat))
else:
print(feat)
if feat is None:
break
print('end LoadBatch')
def input_feat_and_label(self):
feat, label, length = self.kaldi_io_nstream.LoadNextNstreams()
if length is None:
return False
if len(label) != self.nnet_conf.batch_size:
return False
sparse_label = sparse_tuple_from(label)
self.input_queue.put((feat, sparse_label, length))
self.num_batch_total += 1
for i in length:
self.num_frames_total += i
print('total_batch_num**********', self.num_batch_total, '***********')
return True
def input_feat_and_label(self):
strat_io_time = time.time()
feat, label, length = self.kaldi_io_nstream.LoadNextNstreams()
end_io_time = time.time()
# print('*************io time**********************',end_io_time-strat_io_time)
if length is None:
return False
if len(label) != self.nnet_conf.batch_size:
return False
sparse_label = sparse_tuple_from(label)
self.input_queue.put((feat, sparse_label, length))
self.num_batch_total += 1
print('total_batch_num**********', self.num_batch_total, '***********')
return True
def InputFeat(self, input_lock):
while True:
input_lock.acquire()
'''
if 'ctc' in self.criterion_cf or 'whole' in self.criterion_cf:
if 'cnn' in self.criterion_cf:
feat,label,length = self.kaldi_io_nstream.CnnLoadNextNstreams()
else:
feat,label,length = self.kaldi_io_nstream.WholeLoadNextNstreams()
if length is None:
break
print(np.shape(feat),np.shape(label), np.shape(length))
if len(label) != self.batch_size_cf:
break
if 'ctc' in self.criterion_cf:
sparse_label = sparse_tuple_from(label)
self.input_queue.put((feat,sparse_label,length))
else:
self.input_queue.put((feat,label,length))
elif 'ce' in self.criterion_cf:
if 'cnn' in self.criterion_cf:
feat_array, label_array, length_array = self.kaldi_io_nstream.CnnSliceLoadNextNstreams()
else:
feat_array, label_array, length_array = self.kaldi_io_nstream.SliceLoadNextNstreams()
if length_array is None:
break
print(np.shape(feat_array),np.shape(label_array), np.shape(length_array))
if len(label_array[0]) != self.batch_size_cf:
break
self.input_queue.put((feat_array, label_array, length_array))
'''
feat,label,length = self.kaldi_io_nstream.LoadBatch()
if length is None:
break
print(np.shape(feat),np.shape(label), np.shape(length))
sys.stdout.flush()
if 'ctc' in self.criterion_cf:
sparse_label = sparse_tuple_from(label)
self.input_queue.put((feat,sparse_label,length))
else:
self.input_queue.put((feat,label,length))
self.num_batch_total += 1
# if self.num_batch_total % 3000 == 0:
# self.SaveModel()
# self.AdjustLearnRate()
print('total_batch_num**********',self.num_batch_total,'***********')
input_lock.release()
self.input_queue.put((None, None, None))
def InputFeat(self, input_lock):
while True:
input_lock.acquire()
feat,label,length = self.kaldi_io_nstream.LoadNextNstreams()
if length is None:
break
if len(label) != self.batch_size_cf:
break
sparse_label = sparse_tuple_from(label)
self.input_queue.put((feat,sparse_label,length))
self.num_batch_total += 1
if self.num_batch_total % 3000 == 0:
self.SaveModel()
self.AdjustLearnRate()
print('total_batch_num**********',self.num_batch_total,'***********')
input_lock.release()
self.input_queue.put((None, None, None))