def _classify_relations(self, entity_spans, size_embeddings, relations,
rel_masks, h, chunk_start):
batch_size = relations.shape[0]
# create chunks if necessary
if relations.shape[1] > self._max_pairs:
relations = relations[:, chunk_start:chunk_start + self._max_pairs]
rel_masks = rel_masks[:, chunk_start:chunk_start + self._max_pairs]
h = h[:, :relations.shape[1], :]
# get pairs of entity candidate representations
entity_pairs = util.batch_index(entity_spans, relations)
entity_pairs = entity_pairs.view(batch_size, entity_pairs.shape[1], -1)
# get corresponding size embeddings
size_pair_embeddings = util.batch_index(size_embeddings, relations)
size_pair_embeddings = size_pair_embeddings.view(
batch_size, size_pair_embeddings.shape[1], -1)
# relation context (context between entity candidate pair)
rel_ctx = rel_masks * h
rel_ctx = rel_ctx.max(dim=2)[0]
# create relation candidate representations including context, max pooled entity candidate pairs
# and corresponding size embeddings
rel_repr = torch.cat([rel_ctx, entity_pairs, size_pair_embeddings],
dim=2)
rel_repr = self.dropout(rel_repr)
# classify relation candidates
chunk_rel_logits = self.rel_classifier(rel_repr)
return chunk_rel_logits
def _classify_relations(self, entity_spans, size_embeddings, relations, rel_masks, h, chunk_start):
batch_size = relations.shape[0]
# create chunks if necessary
if relations.shape[1] > self._max_pairs:
relations = relations[:, chunk_start:chunk_start + self._max_pairs]
rel_masks = rel_masks[:, chunk_start:chunk_start + self._max_pairs]
h = h[:, :relations.shape[1], :]
# get pairs of entity candidate representations
entity_pairs = util.batch_index(entity_spans, relations)
entity_pairs = entity_pairs.view(batch_size, entity_pairs.shape[1], -1)
# get corresponding size embeddings
size_pair_embeddings = util.batch_index(size_embeddings, relations)
size_pair_embeddings = size_pair_embeddings.view(batch_size, size_pair_embeddings.shape[1], -1)
# relation context (context between entity candidate pair)
# mask non entity candidate tokens
m = ((rel_masks == 0).float() * (-1e30)).unsqueeze(-1)
rel_ctx = m + h
# max pooling
rel_ctx = rel_ctx.max(dim=2)[0]
# set the context vector of neighboring or adjacent entity candidates to zero
rel_ctx[rel_masks.to(torch.uint8).any(-1) == 0] = 0
# create relation candidate representations including context, max pooled entity candidate pairs
# and corresponding size embeddings
rel_repr = torch.cat([rel_ctx, entity_pairs, size_pair_embeddings], dim=2)
rel_repr = self.dropout(rel_repr)
# classify relation candidates
chunk_rel_logits = self.rel_classifier(rel_repr)
return chunk_rel_logits