def __call__(self, xs, ilens):
"""RNNP forward
:param xs:
:param ilens:
:return:
"""
logging.info(self.__class__.__name__ + ' input lengths: ' + str(ilens))
for layer in six.moves.range(self.elayers):
if "lstm" in self.typ:
_, _, ys = self[self.rnn_label + str(layer)](None, None, xs)
else:
_, ys = self[self.rnn_label + str(layer)](None, xs)
# ys: utt list of frame x cdim x 2 (2: means bidirectional)
# TODO(watanabe) replace subsample and FC layer with CNN
ys, ilens = _subsamplex(ys, self.subsample[layer + 1])
# (sum _utt frame_utt) x dim
ys = self['bt' + str(layer)](F.vstack(ys))
xs = F.split_axis(ys, np.cumsum(ilens[:-1]), axis=0)
# final tanh operation
xs = F.split_axis(F.tanh(F.vstack(xs)), np.cumsum(ilens[:-1]), axis=0)
# 1 utterance case, it becomes an array, so need to make a utt tuple
if not isinstance(xs, tuple):
xs = [xs]
return xs, ilens # x: utt list of frame x dim
def __call__(self, xs, ilens):
"""RNNP forward.
Args:
xs (chainer.Variable): Batch of padded charactor ids. (B, Tmax)
ilens (chainer.Variable): Batch of length of each input batch. (B,)
Returns:
xs (chainer.Variable):subsampled vector of xs.
chainer.Variable: Subsampled vector of ilens.
"""
logging.info(self.__class__.__name__ + " input lengths: " + str(ilens))
for layer in six.moves.range(self.elayers):
if "lstm" in self.typ:
_, _, ys = self[self.rnn_label + str(layer)](None, None, xs)
else:
_, ys = self[self.rnn_label + str(layer)](None, xs)
# ys: utt list of frame x cdim x 2 (2: means bidirectional)
# TODO(watanabe) replace subsample and FC layer with CNN
ys, ilens = _subsamplex(ys, self.subsample[layer + 1])
# (sum _utt frame_utt) x dim
ys = self["bt" + str(layer)](F.vstack(ys))
xs = F.split_axis(ys, np.cumsum(ilens[:-1]), axis=0)
# final tanh operation
xs = F.split_axis(F.tanh(F.vstack(xs)), np.cumsum(ilens[:-1]), axis=0)
# 1 utterance case, it becomes an array, so need to make a utt tuple
if not isinstance(xs, tuple):
xs = [xs]
return xs, ilens # x: utt list of frame x dim