def unwrap_sentence(self, sentences, transitions):
# Build Tokens
x = sentences
# Build Transitions
t = transitions
example = Example()
example.tokens = to_gpu(Variable(torch.from_numpy(x), volatile=not self.training))
example.transitions = t
return example
def build_example(self, sentences, transitions):
batch_size = sentences.shape[0]
# Build Tokens
x = sentences
# Build Transitions
t = transitions
example = Example()
example.tokens = to_gpu(Variable(torch.from_numpy(x), volatile=not self.training))
example.transitions = t
return example
def unwrap_sentence_pair(self, sentences, transitions):
# Build Tokens
x_prem = sentences[:, :, 0]
x_hyp = sentences[:, :, 1]
x = np.concatenate([x_prem, x_hyp], axis=0)
# Build Transitions
t_prem = transitions[:, :, 0]
t_hyp = transitions[:, :, 1]
t = np.concatenate([t_prem, t_hyp], axis=0)
example = Example()
example.tokens = to_gpu(Variable(torch.from_numpy(x), volatile=not self.training))
example.transitions = t
return example
def forward(self, sentences, transitions, y_batch=None, **kwargs):
# Useful when investigating dynamic batching:
# self.seq_lengths = sentences.shape[1] - (sentences == 0).sum(1)
x = self.unwrap(sentences, transitions)
emb = self.run_embed(x)
if self.data_type=="mt":
hn, cn, output=self.run_rnn(emb)
example=Example()
example.bufs=emb
example.tokens=x
return example, hn, output, None, None, None
hh = torch.squeeze(self.run_rnn(emb))
h = self.wrap(hh)
output = self.mlp(self.build_features(h))
return output
def build_example(self, sentences, transitions):
batch_size = sentences.shape[0]
# sentences: (#batches, #feature, #2)
# Build Tokens
x_prem = sentences[:,:,0]
x_hyp = sentences[:,:,1]
x = np.concatenate([x_prem, x_hyp], axis=0)
# Build Transitions
t_prem = transitions[:,:,0]
t_hyp = transitions[:,:,1]
t = np.concatenate([t_prem, t_hyp], axis=0)
example = Example()
example.tokens = to_gpu(Variable(torch.from_numpy(x), volatile=not self.training))
example.transitions = t
return example
