def forward(self,
words: torch.Tensor,
word_lengths: torch.Tensor,
word_indices: torch.Tensor,
word_idxs: torch.Tensor,
sent_context: torch.Tensor,
labs: torch.Tensor,
lens: torch.Tensor = None) -> torch.Tensor:
# pylint: disable=arguments-differ,not-callable
"""
Compute the negative of the log-likelihood.
Parameters
----------
words : torch.Tensor
Word tensors ``[sent_length x max_word_lenth x n_chars]``.
word_lengths : torch.Tensor
Contains the length of each word (in characters), with shape
``[sent_length]``.
word_indices : torch.Tensor
Contains sorted indices of words by length, with shape
``[sent_length]``.
word_idxs : torch.Tensor
Word indices with shape ``[sent_length]``.
sent_context : torch.Tensor
Sentence context.
labs : torch.Tensor
Corresponding target label sequence with shape ``[sent_length]``.
lens : torch.Tensor, optional (default: None)
Gives the length of each sentence in the batch ``[batch_size]``.
Returns
-------
torch.Tensor
The negative log-likelihood evaluated at the inputs.
"""
if lens is None:
lens = torch.tensor([words.size(0)], device=words.device)
mask = sequence_mask(lens)
# Fake batch dimension for ``labs``.
labs = labs.unsqueeze(0)
# labs: ``[1 x sent_length]``
# Gather word feats.
feats = self._feats(words, word_lengths, word_indices, word_idxs,
sent_context)
# feats: ``[1 x sent_length x n_labels]``
loglik = self.crf(feats, labs, mask=mask)
return -1. * loglik
def predict(self,
words: torch.Tensor,
word_lengths: torch.Tensor,
word_indices: torch.Tensor,
word_idxs: torch.Tensor,
sent_context: torch.Tensor,
lens: torch.Tensor = None) -> List[Tuple[List[int], float]]:
# pylint: disable=not-callable
"""
Compute the best tag sequence.
Parameters
----------
words : torch.Tensor
Word tensors ``[sent_length x max_word_lenth x n_chars]``.
word_lengths : torch.Tensor
Contains the length of each word (in characters), with shape
``[sent_length]``.
word_indices : torch.Tensor
Contains sorted indices of words by length, with shape
``[sent_length]``.
word_idxs : torch.Tensor
Word indices with shape ``[sent_length]``.
sent_context : torch.Tensor
Sentence context.
lens : torch.Tensor, optional (default: None)
Gives the length of each sentence in the batch ``[batch_size]``.
Returns
-------
List[List[int]]
The best path for each sentence in the batch.
"""
if lens is None:
lens = torch.tensor([words.size(0)], device=words.device)
mask = sequence_mask(lens)
# Gather word feats.
feats = self._feats(words, word_lengths, word_indices, word_idxs,
sent_context)
# feats: ``[1 x sent_length x n_labels]``
# Run features through Viterbi decode algorithm.
preds = self.crf.viterbi_tags(feats, mask)
return preds
def test_sequence_mask(inp, chk, max_len):
"""Test `sequence_mask()` method."""
res = utils.sequence_mask(inp, max_len=max_len)
utils.assert_equal(res, chk)