def forward(
self, # type: ignore
token_ids: torch.LongTensor,
type_ids: torch.LongTensor,
offsets: torch.LongTensor,
wordpiece_mask: torch.BoolTensor,
dep_idxs: torch.LongTensor,
dep_tags: torch.LongTensor,
pos_tags: torch.LongTensor,
word_mask: torch.BoolTensor,
):
embedded_text_input = self.get_word_embedding(
token_ids=token_ids,
offsets=offsets,
wordpiece_mask=wordpiece_mask,
type_ids=type_ids,
)
if self.pos_embedding is not None:
embedded_pos_tags = self.pos_embedding(pos_tags)
embedded_text_input = torch.cat(
[embedded_text_input, embedded_pos_tags], -1)
if self.fuse_layer is not None:
embedded_text_input = self.fuse_layer(embedded_text_input)
# todo compare normal dropout with InputVariationalDropout
embedded_text_input = self._input_dropout(embedded_text_input)
if self.additional_encoder is not None:
if self.config.additional_layer_type == "transformer":
extended_attention_mask = self.bert.get_extended_attention_mask(
word_mask, word_mask.size(), word_mask.device)
encoded_text = self.additional_encoder(
hidden_states=embedded_text_input,
attention_mask=extended_attention_mask)[0]
else:
encoded_text = self.additional_encoder(
inputs=embedded_text_input, mask=word_mask)
else:
encoded_text = embedded_text_input
batch_size, _, encoding_dim = encoded_text.size()
head_sentinel = self._head_sentinel.expand(batch_size, 1, encoding_dim)
# Concatenate the head sentinel onto the sentence representation.
encoded_text = torch.cat([head_sentinel, encoded_text], 1)
word_mask = torch.cat([word_mask.new_ones(batch_size, 1), word_mask],
1)
dep_idxs = torch.cat([dep_idxs.new_zeros(batch_size, 1), dep_idxs], 1)
dep_tags = torch.cat([dep_tags.new_zeros(batch_size, 1), dep_tags], 1)
encoded_text = self._dropout(encoded_text)
# shape (batch_size, sequence_length, arc_representation_dim)
head_arc_representation = self._dropout(
self.head_arc_feedforward(encoded_text))
child_arc_representation = self._dropout(
self.child_arc_feedforward(encoded_text))
# shape (batch_size, sequence_length, tag_representation_dim)
head_tag_representation = self._dropout(
self.head_tag_feedforward(encoded_text))
child_tag_representation = self._dropout(
self.child_tag_feedforward(encoded_text))
# shape (batch_size, sequence_length, sequence_length)
attended_arcs = self.arc_attention(head_arc_representation,
child_arc_representation)
minus_inf = -1e8
minus_mask = ~word_mask * minus_inf
attended_arcs = attended_arcs + minus_mask.unsqueeze(
2) + minus_mask.unsqueeze(1)
if self.training:
predicted_heads, predicted_head_tags = self._greedy_decode(
head_tag_representation, child_tag_representation,
attended_arcs, word_mask)
else:
predicted_heads, predicted_head_tags = self._mst_decode(
head_tag_representation, child_tag_representation,
attended_arcs, word_mask)
arc_nll, tag_nll = self._construct_loss(
head_tag_representation=head_tag_representation,
child_tag_representation=child_tag_representation,
attended_arcs=attended_arcs,
head_indices=dep_idxs,
head_tags=dep_tags,
mask=word_mask,
)
return predicted_heads, predicted_head_tags, arc_nll, tag_nll
def _parse(
self,
embedded_text_input: torch.Tensor,
mask: torch.BoolTensor,
head_tags: torch.LongTensor = None,
head_indices: torch.LongTensor = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor,
torch.Tensor]:
embedded_text_input = self._input_dropout(embedded_text_input)
encoded_text = self.encoder(embedded_text_input, mask)
batch_size, _, encoding_dim = encoded_text.size()
head_sentinel = self._head_sentinel.expand(batch_size, 1, encoding_dim)
# Concatenate the head sentinel onto the sentence representation.
encoded_text = torch.cat([head_sentinel, encoded_text], 1)
mask = torch.cat([mask.new_ones(batch_size, 1), mask], 1)
if head_indices is not None:
head_indices = torch.cat(
[head_indices.new_zeros(batch_size, 1), head_indices], 1)
if head_tags is not None:
head_tags = torch.cat(
[head_tags.new_zeros(batch_size, 1), head_tags], 1)
encoded_text = self._dropout(encoded_text)
# shape (batch_size, sequence_length, arc_representation_dim)
head_arc_representation = self._dropout(
self.head_arc_feedforward(encoded_text))
child_arc_representation = self._dropout(
self.child_arc_feedforward(encoded_text))
# shape (batch_size, sequence_length, tag_representation_dim)
head_tag_representation = self._dropout(
self.head_tag_feedforward(encoded_text))
child_tag_representation = self._dropout(
self.child_tag_feedforward(encoded_text))
# shape (batch_size, sequence_length, sequence_length)
attended_arcs = self.arc_attention(head_arc_representation,
child_arc_representation)
minus_inf = -1e8
minus_mask = ~mask * minus_inf
attended_arcs = attended_arcs + minus_mask.unsqueeze(
2) + minus_mask.unsqueeze(1)
if self.training or not self.use_mst_decoding_for_validation:
predicted_heads, predicted_head_tags = self._greedy_decode(
head_tag_representation, child_tag_representation,
attended_arcs, mask)
else:
predicted_heads, predicted_head_tags = self._mst_decode(
head_tag_representation, child_tag_representation,
attended_arcs, mask)
if head_indices is not None and head_tags is not None:
arc_nll, tag_nll = self._construct_loss(
head_tag_representation=head_tag_representation,
child_tag_representation=child_tag_representation,
attended_arcs=attended_arcs,
head_indices=head_indices,
head_tags=head_tags,
mask=mask,
)
else:
arc_nll, tag_nll = self._construct_loss(
head_tag_representation=head_tag_representation,
child_tag_representation=child_tag_representation,
attended_arcs=attended_arcs,
head_indices=predicted_heads.long(),
head_tags=predicted_head_tags.long(),
mask=mask,
)
return predicted_heads, predicted_head_tags, mask, arc_nll, tag_nll
def _parse(
self,
embedded_text_input: torch.Tensor,
mask: torch.BoolTensor,
head_tags: torch.LongTensor = None,
head_indices: torch.LongTensor = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
embedded_text_input = self._input_dropout(embedded_text_input)
encoded_text = self.encoder(embedded_text_input, mask)
batch_size, sequence_length, encoding_dim = encoded_text.size()
head_sentinel = self._head_sentinel.expand(batch_size, 1, encoding_dim)
# Concatenate the head sentinel onto the sentence representation.
encoded_text = torch.cat([head_sentinel, encoded_text], 1)
mask = torch.cat([mask.new_ones(batch_size, 1), mask], 1)
if head_indices is not None:
head_indices = torch.cat([head_indices.new_zeros(batch_size, 1), head_indices], 1)
if head_tags is not None:
head_tags = torch.cat([head_tags.new_zeros(batch_size, 1), head_tags], 1)
encoded_text = self._dropout(encoded_text)
# shape (batch_size, sequence_length, arc_representation_dim)
head_arc_representation = self.head_arc_feedforward(encoded_text)
child_arc_representation = self.child_arc_feedforward(encoded_text)
# shape (batch_size, sequence_length, tag_representation_dim)
head_tag_representation = self.head_tag_feedforward(encoded_text)
child_tag_representation = self.child_tag_feedforward(encoded_text)
# calculate dimensions again as sequence_length is now + 1 from adding the head_sentinel
batch_size, sequence_length, arc_dim = head_arc_representation.size()
# now repeat the token representations to form a matrix:
# shape (batch_size, sequence_length, sequence_length, arc_representation_dim)
heads = head_arc_representation.repeat(1, sequence_length, 1).reshape(batch_size, sequence_length, sequence_length, arc_dim) # heads in one direction
deps = child_arc_representation.repeat(1, sequence_length, 1).reshape(batch_size, sequence_length, sequence_length, arc_dim).transpose(1, 2) # deps in the other direction
# shape (batch_size, sequence_length, sequence_length, arc_representation_dim)
combined_arcs = self.activation(heads + deps)
# shape (batch_size, sequence_length, sequence_length)
attended_arcs = self.arc_out_layer(combined_arcs).squeeze(3)
minus_inf = -1e8
minus_mask = ~mask * minus_inf
attended_arcs = attended_arcs + minus_mask.unsqueeze(2) + minus_mask.unsqueeze(1)
if self.training or not self.use_mst_decoding_for_validation:
predicted_heads, predicted_head_tags = self._greedy_decode(
head_tag_representation, child_tag_representation, attended_arcs, mask
)
else:
predicted_heads, predicted_head_tags = self._mst_decode(
head_tag_representation, child_tag_representation, attended_arcs, mask
)
if head_indices is not None and head_tags is not None:
arc_nll, tag_nll = self._construct_loss(
head_tag_representation=head_tag_representation,
child_tag_representation=child_tag_representation,
attended_arcs=attended_arcs,
head_indices=head_indices,
head_tags=head_tags,
mask=mask,
)
else:
arc_nll, tag_nll = self._construct_loss(
head_tag_representation=head_tag_representation,
child_tag_representation=child_tag_representation,
attended_arcs=attended_arcs,
head_indices=predicted_heads.long(),
head_tags=predicted_head_tags.long(),
mask=mask,
)
return predicted_heads, predicted_head_tags, mask, arc_nll, tag_nll
