def forward(self, ent, ent_mask, ent_len, rel, rel_mask, graphs):
device = ent.device
ent_mask = (ent_mask == 0) # reverse mask
rel_mask = (rel_mask == 0)
init_h = []
for i in range(graphs.batch_size):
init_h.append(ent[i][ent_mask[i]])
init_h.append(rel[i][rel_mask[i]])
init_h = torch.cat(init_h, 0)
feats = init_h
if graphs.number_of_nodes() != len(init_h):
print('Err', graphs.number_of_nodes(), len(init_h), ent_mask,
rel_mask)
else:
for i in range(self.prop):
feats = self.gat[i](graphs, feats)
g_root = feats.index_select(
0,
graphs.filter_nodes(
lambda x: x.data['type'] == NODE_TYPE['root']).to(device))
g_ent = pad(feats.index_select(
0,
graphs.filter_nodes(lambda x: x.data['type'] == NODE_TYPE['entity']
).to(device)).split(ent_len),
out_type='tensor')
return g_ent, g_root
def get_sentiment():
text = request.get_json(force=True)['text']
words = text_to_word_sequence(text)
word_indices = [
words_to_indices[word] + data.offset if word in words_to_indices
and words_to_indices[word] < data.top_words else data.oov_index
for word in words
]
sequence = data.pad([[data.start_index] + word_indices])
with graph.as_default():
sentiment = model.predict(sequence)
return '{0:.2f}% positive'.format(sentiment[0][0] * 100)
def forward(self, inp, mask, ent_len=None):
inp = self.drop(inp)
lens = (mask == 0).sum(-1).long().tolist()
pad_seq = pack_padded_sequence(inp,
lens,
batch_first=True,
enforce_sorted=False)
y, (_h, _c) = self.bilstm(pad_seq)
_h = _h.transpose(0, 1).contiguous()
_h = _h[:, -2:].view(_h.size(0), -1) # two directions of the top-layer
ret = pad(_h.split(ent_len), out_type='tensor')
return ret
def forward(self, inp, mask, ent_len=None):
inp = self.drop(inp) # (tot_ent, max_ent_len, dim_emb)
lens = (mask == 0).sum(-1).long().tolist()
pad_seq = pack_padded_sequence(
inp, lens, batch_first=True,
enforce_sorted=False) # data of shape (sum(ent_len_ij), dim_emb)
# y: packed sequence (dim=dim_h*2), h_t of the last layer for each t
y, (_h, _c) = self.bilstm(pad_seq)
# (tot_ent, num_layers*num_directions, dim_h), hidden state for t=seq_len
_h = _h.transpose(0, 1).contiguous()
# two directions of the top-layer (tot_ent, dim_h*2=d)
_h = _h[:, -2:].view(_h.size(0), -1)
# _h.split: list of len bs, each element is a (num_ent_i, d) tensor
ret = pad(_h.split(ent_len), out_type="tensor") # (bs, max_num_ent, d)
return ret