def __call__(self, hs, ys):
"""Core function of the Warp-CTC layer.
Args:
hs (iterable of chainer.Variable | N-dimention array): Input variable from encoder.
ys (iterable of chainer.Variable | N-dimension array): Input variable of decoder.
Returns:
chainer.Variable: A variable holding a scalar value of the CTC loss.
"""
self.loss = None
ilens = [x.shape[0] for x in hs]
olens = [x.shape[0] for x in ys]
# zero padding for hs
y_hat = linear_tensor(self.ctc_lo, F.dropout(
F.pad_sequence(hs), ratio=self.dropout_rate))
y_hat = F.transpose(y_hat, (1, 0, 2)) # batch x frames x hdim
# get length info
logging.info(self.__class__.__name__ + ' input lengths: ' + str(ilens))
logging.info(self.__class__.__name__ + ' output lengths: ' + str(olens))
# get ctc loss
from chainer_ctc.warpctc import ctc as warp_ctc
self.loss = warp_ctc(y_hat, ilens, [cuda.to_cpu(l.data) for l in ys])[0]
logging.info('ctc loss:' + str(self.loss.data))
return self.loss
def __call__(self, hs, ys):
'''CTC forward
:param hs:
:param ys:
:return:
'''
self.loss = None
ilens = [x.shape[0] for x in hs]
olens = [x.shape[0] for x in ys]
# zero padding for hs
y_hat = linear_tensor(self.ctc_lo, F.dropout(
F.pad_sequence(hs), ratio=self.dropout_rate))
y_hat = F.transpose(y_hat, (1, 0, 2)) # batch x frames x hdim
# get length info
logging.info(self.__class__.__name__ + ' input lengths: ' + str(ilens))
logging.info(self.__class__.__name__ + ' output lengths: ' + str(olens))
# get ctc loss
self.loss = warp_ctc(y_hat, ilens, [cuda.to_cpu(l.data) for l in ys])[0]
logging.info('ctc loss:' + str(self.loss.data))
return self.loss
