def decode(
self, output_dict: Dict[str,
torch.Tensor]) -> Dict[str, torch.Tensor]:
"""
Does constrained viterbi decoding on class probabilities output in :func:`forward`. The
constraint simply specifies that the output tags must be a valid BIO sequence. We add a
``"tags"`` key to the dictionary with the result.
"""
all_predictions = output_dict['class_probabilities']
sequence_lengths = get_lengths_from_binary_sequence_mask(
output_dict["mask"]).data.tolist()
if all_predictions.dim() == 3:
predictions_list = [
all_predictions[i].detach().cpu()
for i in range(all_predictions.size(0))
]
else:
predictions_list = [all_predictions]
all_tags = []
transition_matrix = self.get_viterbi_pairwise_potentials()
for predictions, length in zip(predictions_list, sequence_lengths):
max_likelihood_sequence, _ = viterbi_decode(
predictions[:length], transition_matrix)
tags = [
self.vocab.get_token_from_index(x, namespace="labels")
for x in max_likelihood_sequence
]
all_tags.append(tags)
output_dict['tags'] = all_tags
return output_dict
def test_forward_pulls_out_correct_tensor_with_sequence_lengths(self):
lstm = LSTM(bidirectional=True,
num_layers=3,
input_size=3,
hidden_size=7,
batch_first=True)
encoder = PytorchSeq2SeqWrapper(lstm)
input_tensor = torch.rand([5, 7, 3])
input_tensor[1, 6:, :] = 0
input_tensor[2, 4:, :] = 0
input_tensor[3, 2:, :] = 0
input_tensor[4, 1:, :] = 0
mask = torch.ones(5, 7)
mask[1, 6:] = 0
mask[2, 4:] = 0
mask[3, 2:] = 0
mask[4, 1:] = 0
sequence_lengths = get_lengths_from_binary_sequence_mask(mask)
packed_sequence = pack_padded_sequence(input_tensor,
sequence_lengths.data.tolist(),
batch_first=True)
lstm_output, _ = lstm(packed_sequence)
encoder_output = encoder(input_tensor, mask)
lstm_tensor, _ = pad_packed_sequence(lstm_output, batch_first=True)
assert_almost_equal(encoder_output.data.numpy(),
lstm_tensor.data.numpy())
def test_forward_pulls_out_correct_tensor_with_unsorted_batches(self):
lstm = LSTM(bidirectional=True, num_layers=3, input_size=3, hidden_size=7, batch_first=True)
encoder = PytorchSeq2VecWrapper(lstm)
input_tensor = torch.rand([5, 7, 3])
input_tensor[0, 3:, :] = 0
input_tensor[1, 4:, :] = 0
input_tensor[2, 2:, :] = 0
input_tensor[3, 6:, :] = 0
mask = torch.ones(5, 7)
mask[0, 3:] = 0
mask[1, 4:] = 0
mask[2, 2:] = 0
mask[3, 6:] = 0
sequence_lengths = get_lengths_from_binary_sequence_mask(mask)
sorted_inputs, sorted_sequence_lengths, restoration_indices, _ = sort_batch_by_length(input_tensor,
sequence_lengths)
packed_sequence = pack_padded_sequence(sorted_inputs,
sorted_sequence_lengths.tolist(),
batch_first=True)
_, state = lstm(packed_sequence)
# Transpose output state, extract the last forward and backward states and
# reshape to be of dimension (batch_size, 2 * hidden_size).
sorted_transposed_state = state[0].transpose(0, 1).index_select(0, restoration_indices)
reshaped_state = sorted_transposed_state[:, -2:, :].contiguous()
explicitly_concatenated_state = torch.cat([reshaped_state[:, 0, :].squeeze(1),
reshaped_state[:, 1, :].squeeze(1)], -1)
encoder_output = encoder(input_tensor, mask)
assert_almost_equal(encoder_output.data.numpy(), explicitly_concatenated_state.data.numpy())
def test_forward_pulls_out_correct_tensor_for_unsorted_batches(self):
lstm = LSTM(bidirectional=True,
num_layers=3,
input_size=3,
hidden_size=7,
batch_first=True)
encoder = PytorchSeq2SeqWrapper(lstm)
input_tensor = torch.rand([5, 7, 3])
input_tensor[0, 3:, :] = 0
input_tensor[1, 4:, :] = 0
input_tensor[2, 2:, :] = 0
input_tensor[3, 6:, :] = 0
mask = torch.ones(5, 7)
mask[0, 3:] = 0
mask[1, 4:] = 0
mask[2, 2:] = 0
mask[3, 6:] = 0
sequence_lengths = get_lengths_from_binary_sequence_mask(mask)
sorted_inputs, sorted_sequence_lengths, restoration_indices, _ = sort_batch_by_length(
input_tensor, sequence_lengths)
packed_sequence = pack_padded_sequence(
sorted_inputs,
sorted_sequence_lengths.data.tolist(),
batch_first=True)
lstm_output, _ = lstm(packed_sequence)
encoder_output = encoder(input_tensor, mask)
lstm_tensor, _ = pad_packed_sequence(lstm_output, batch_first=True)
assert_almost_equal(
encoder_output.data.numpy(),
lstm_tensor.index_select(0, restoration_indices).data.numpy())
def test_get_sequence_lengths_from_binary_mask(self):
binary_mask = torch.ByteTensor([[1, 1, 1, 0, 0, 0], [1, 1, 0, 0, 0, 0],
[1, 1, 1, 1, 1, 1], [1, 0, 0, 0, 0,
0]])
lengths = util.get_lengths_from_binary_sequence_mask(binary_mask)
numpy.testing.assert_array_equal(lengths.numpy(),
numpy.array([3, 2, 6, 1]))
def test_decode_runs_correctly(self):
training_tensors = self.dataset.as_tensor_dict()
output_dict = self.model(**training_tensors)
decode_output_dict = self.model.decode(output_dict)
lengths = get_lengths_from_binary_sequence_mask(
decode_output_dict["mask"]).data.tolist()
# Hard to check anything concrete which we haven't checked in the above
# test, so we'll just check that the tags are equal to the lengths
# of the individual instances, rather than the max length.
for prediction, length in zip(decode_output_dict["tags"], lengths):
assert len(prediction) == length
def test_get_sequence_lengths_converts_to_long_tensor_and_avoids_variable_overflow(
self):
# Tests the following weird behaviour in Pytorch 0.1.12
# doesn't happen for our sequence masks:
#
# mask = torch.ones([260]).byte()
# mask.sum() # equals 260.
# var_mask = t.a.V(mask)
# var_mask.sum() # equals 4, due to 8 bit precision - the sum overflows.
binary_mask = torch.ones(2, 260).byte()
lengths = util.get_lengths_from_binary_sequence_mask(binary_mask)
numpy.testing.assert_array_equal(lengths.data.numpy(),
numpy.array([260, 260]))
def setUp(self):
super(TestEncoderBase, self).setUp()
self.lstm = LSTM(bidirectional=True, num_layers=3, input_size=3, hidden_size=7, batch_first=True)
self.rnn = RNN(bidirectional=True, num_layers=3, input_size=3, hidden_size=7, batch_first=True)
self.encoder_base = _EncoderBase(stateful=True)
tensor = torch.rand([5, 7, 3])
tensor[1, 6:, :] = 0
tensor[3, 2:, :] = 0
self.tensor = tensor
mask = torch.ones(5, 7)
mask[1, 6:] = 0
mask[2, :] = 0 # <= completely masked
mask[3, 2:] = 0
mask[4, :] = 0 # <= completely masked
self.mask = mask
self.batch_size = 5
self.num_valid = 3
sequence_lengths = get_lengths_from_binary_sequence_mask(mask)
_, _, restoration_indices, sorting_indices = sort_batch_by_length(tensor, sequence_lengths)
self.sorting_indices = sorting_indices
self.restoration_indices = restoration_indices
def decode(
self, output_dict: Dict[str,
torch.Tensor]) -> Dict[str, torch.Tensor]:
head_tags = output_dict.pop("head_tags").cpu().detach().numpy()
heads = output_dict.pop("heads").cpu().detach().numpy()
mask = output_dict.pop("mask")
lengths = get_lengths_from_binary_sequence_mask(mask)
head_tag_labels = []
head_indices = []
for instance_heads, instance_tags, length in zip(
heads, head_tags, lengths):
instance_heads = list(instance_heads[1:length])
instance_tags = instance_tags[1:length]
labels = [
self.vocab.get_token_from_index(label, "head_tags")
for label in instance_tags
]
head_tag_labels.append(labels)
head_indices.append(instance_heads)
output_dict["predicted_dependencies"] = head_tag_labels
output_dict["predicted_heads"] = head_indices
return output_dict
def forward(
self,
sequence_tensor: torch.FloatTensor,
span_indices: torch.LongTensor,
sequence_mask: torch.LongTensor = None,
span_indices_mask: torch.LongTensor = None) -> torch.FloatTensor:
# Both of shape (batch_size, sequence_length, embedding_size / 2)
forward_sequence, backward_sequence = sequence_tensor.split(int(
self._input_dim / 2),
dim=-1)
forward_sequence = forward_sequence.contiguous()
backward_sequence = backward_sequence.contiguous()
# shape (batch_size, num_spans)
span_starts, span_ends = [
index.squeeze(-1) for index in span_indices.split(1, dim=-1)
]
if span_indices_mask is not None:
span_starts = span_starts * span_indices_mask
span_ends = span_ends * span_indices_mask
# We want `exclusive` span starts, so we remove 1 from the forward span starts
# as the AllenNLP ``SpanField`` is inclusive.
# shape (batch_size, num_spans)
exclusive_span_starts = span_starts - 1
# shape (batch_size, num_spans, 1)
start_sentinel_mask = (
exclusive_span_starts == -1).long().unsqueeze(-1)
# We want `exclusive` span ends for the backward direction
# (so that the `start` of the span in that direction is exlusive), so
# we add 1 to the span ends as the AllenNLP ``SpanField`` is inclusive.
exclusive_span_ends = span_ends + 1
if sequence_mask is not None:
# shape (batch_size)
sequence_lengths = util.get_lengths_from_binary_sequence_mask(
sequence_mask)
else:
# shape (batch_size), filled with the sequence length size of the sequence_tensor.
sequence_lengths = (
torch.ones_like(sequence_tensor[:, 0, 0], dtype=torch.long) *
sequence_tensor.size(1))
# shape (batch_size, num_spans, 1)
end_sentinel_mask = (exclusive_span_ends == sequence_lengths.unsqueeze(
-1)).long().unsqueeze(-1)
# As we added 1 to the span_ends to make them exclusive, which might have caused indices
# equal to the sequence_length to become out of bounds, we multiply by the inverse of the
# end_sentinel mask to erase these indices (as we will replace them anyway in the block below).
# The same argument follows for the exclusive span start indices.
exclusive_span_ends = exclusive_span_ends * (
1 - end_sentinel_mask.squeeze(-1))
exclusive_span_starts = exclusive_span_starts * (
1 - start_sentinel_mask.squeeze(-1))
# We'll check the indices here at runtime, because it's difficult to debug
# if this goes wrong and it's tricky to get right.
if (exclusive_span_starts < 0).any() or (
exclusive_span_ends > sequence_lengths.unsqueeze(-1)).any():
raise ValueError(
f"Adjusted span indices must lie inside the length of the sequence tensor, "
f"but found: exclusive_span_starts: {exclusive_span_starts}, "
f"exclusive_span_ends: {exclusive_span_ends} for a sequence tensor with lengths "
f"{sequence_lengths}.")
# Forward Direction: start indices are exclusive. Shape (batch_size, num_spans, input_size / 2)
forward_start_embeddings = util.batched_index_select(
forward_sequence, exclusive_span_starts)
# Forward Direction: end indices are inclusive, so we can just use span_ends.
# Shape (batch_size, num_spans, input_size / 2)
forward_end_embeddings = util.batched_index_select(
forward_sequence, span_ends)
# Backward Direction: The backward start embeddings use the `forward` end
# indices, because we are going backwards.
# Shape (batch_size, num_spans, input_size / 2)
backward_start_embeddings = util.batched_index_select(
backward_sequence, exclusive_span_ends)
# Backward Direction: The backward end embeddings use the `forward` start
# indices, because we are going backwards.
# Shape (batch_size, num_spans, input_size / 2)
backward_end_embeddings = util.batched_index_select(
backward_sequence, span_starts)
if self._use_sentinels:
# If we're using sentinels, we need to replace all the elements which were
# outside the dimensions of the sequence_tensor with either the start sentinel,
# or the end sentinel.
float_end_sentinel_mask = end_sentinel_mask.float()
float_start_sentinel_mask = start_sentinel_mask.float()
forward_start_embeddings = forward_start_embeddings * (1 - float_start_sentinel_mask) \
+ float_start_sentinel_mask * self._start_sentinel
backward_start_embeddings = backward_start_embeddings * (1 - float_end_sentinel_mask) \
+ float_end_sentinel_mask * self._end_sentinel
# Now we combine the forward and backward spans in the manner specified by the
# respective combinations and concatenate these representations.
# Shape (batch_size, num_spans, forward_combination_dim)
forward_spans = util.combine_tensors(
self._forward_combination,
[forward_start_embeddings, forward_end_embeddings])
# Shape (batch_size, num_spans, backward_combination_dim)
backward_spans = util.combine_tensors(
self._backward_combination,
[backward_start_embeddings, backward_end_embeddings])
# Shape (batch_size, num_spans, forward_combination_dim + backward_combination_dim)
span_embeddings = torch.cat([forward_spans, backward_spans], -1)
if self._span_width_embedding is not None:
# Embed the span widths and concatenate to the rest of the representations.
if self._bucket_widths:
span_widths = util.bucket_values(
span_ends - span_starts,
num_total_buckets=self._num_width_embeddings)
else:
span_widths = span_ends - span_starts
span_width_embeddings = self._span_width_embedding(span_widths)
return torch.cat([span_embeddings, span_width_embeddings], -1)
if span_indices_mask is not None:
return span_embeddings * span_indices_mask.float().unsqueeze(-1)
return span_embeddings
def forward(self,
context_1: torch.Tensor,
mask_1: torch.Tensor,
context_2: torch.Tensor,
mask_2: torch.Tensor) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
# pylint: disable=arguments-differ
"""
Given the forward (or backward) representations of sentence1 and sentence2, apply four bilateral
matching functions between them in one direction.
Parameters
----------
context_1 : ``torch.Tensor``
Tensor of shape (batch_size, seq_len1, hidden_dim) representing the encoding of the first sentence.
mask_1 : ``torch.Tensor``
Binary Tensor of shape (batch_size, seq_len1), indicating which
positions in the first sentence are padding (0) and which are not (1).
context_2 : ``torch.Tensor``
Tensor of shape (batch_size, seq_len2, hidden_dim) representing the encoding of the second sentence.
mask_2 : ``torch.Tensor``
Binary Tensor of shape (batch_size, seq_len2), indicating which
positions in the second sentence are padding (0) and which are not (1).
Returns
-------
A tuple of matching vectors for the two sentences. Each of which is a list of
matching vectors of shape (batch, seq_len, num_perspectives or 1)
"""
assert (not mask_2.requires_grad) and (not mask_1.requires_grad)
assert context_1.size(-1) == context_2.size(-1) == self.hidden_dim
# (batch,)
len_1 = get_lengths_from_binary_sequence_mask(mask_1)
len_2 = get_lengths_from_binary_sequence_mask(mask_2)
# (batch, seq_len*)
mask_1, mask_2 = mask_1.float(), mask_2.float()
# explicitly set masked weights to zero
# (batch_size, seq_len*, hidden_dim)
context_1 = context_1 * mask_1.unsqueeze(-1)
context_2 = context_2 * mask_2.unsqueeze(-1)
# array to keep the matching vectors for the two sentences
matching_vector_1: List[torch.Tensor] = []
matching_vector_2: List[torch.Tensor] = []
# Step 0. unweighted cosine
# First calculate the cosine similarities between each forward
# (or backward) contextual embedding and every forward (or backward)
# contextual embedding of the other sentence.
# (batch, seq_len1, seq_len2)
cosine_sim = F.cosine_similarity(context_1.unsqueeze(-2), context_2.unsqueeze(-3), dim=3)
# (batch, seq_len*, 1)
cosine_max_1 = masked_max(cosine_sim, mask_2.unsqueeze(-2), dim=2, keepdim=True)
cosine_mean_1 = masked_mean(cosine_sim, mask_2.unsqueeze(-2), dim=2, keepdim=True)
cosine_max_2 = masked_max(cosine_sim.permute(0, 2, 1), mask_1.unsqueeze(-2), dim=2, keepdim=True)
cosine_mean_2 = masked_mean(cosine_sim.permute(0, 2, 1), mask_1.unsqueeze(-2), dim=2, keepdim=True)
matching_vector_1.extend([cosine_max_1, cosine_mean_1])
matching_vector_2.extend([cosine_max_2, cosine_mean_2])
# Step 1. Full-Matching
# Each time step of forward (or backward) contextual embedding of one sentence
# is compared with the last time step of the forward (or backward)
# contextual embedding of the other sentence
if self.with_full_match:
# (batch, 1, hidden_dim)
if self.is_forward:
# (batch, 1, hidden_dim)
last_position_1 = (len_1 - 1).clamp(min=0)
last_position_1 = last_position_1.view(-1, 1, 1).expand(-1, 1, self.hidden_dim)
last_position_2 = (len_2 - 1).clamp(min=0)
last_position_2 = last_position_2.view(-1, 1, 1).expand(-1, 1, self.hidden_dim)
context_1_last = context_1.gather(1, last_position_1)
context_2_last = context_2.gather(1, last_position_2)
else:
context_1_last = context_1[:, 0:1, :]
context_2_last = context_2[:, 0:1, :]
# (batch, seq_len*, num_perspectives)
matching_vector_1_full = multi_perspective_match(context_1,
context_2_last,
self.full_match_weights)
matching_vector_2_full = multi_perspective_match(context_2,
context_1_last,
self.full_match_weights_reversed)
matching_vector_1.extend(matching_vector_1_full)
matching_vector_2.extend(matching_vector_2_full)
# Step 2. Maxpooling-Matching
# Each time step of forward (or backward) contextual embedding of one sentence
# is compared with every time step of the forward (or backward)
# contextual embedding of the other sentence, and only the max value of each
# dimension is retained.
if self.with_maxpool_match:
# (batch, seq_len1, seq_len2, num_perspectives)
matching_vector_max = multi_perspective_match_pairwise(context_1,
context_2,
self.maxpool_match_weights)
# (batch, seq_len*, num_perspectives)
matching_vector_1_max = masked_max(matching_vector_max,
mask_2.unsqueeze(-2).unsqueeze(-1),
dim=2)
matching_vector_1_mean = masked_mean(matching_vector_max,
mask_2.unsqueeze(-2).unsqueeze(-1),
dim=2)
matching_vector_2_max = masked_max(matching_vector_max.permute(0, 2, 1, 3),
mask_1.unsqueeze(-2).unsqueeze(-1),
dim=2)
matching_vector_2_mean = masked_mean(matching_vector_max.permute(0, 2, 1, 3),
mask_1.unsqueeze(-2).unsqueeze(-1),
dim=2)
matching_vector_1.extend([matching_vector_1_max, matching_vector_1_mean])
matching_vector_2.extend([matching_vector_2_max, matching_vector_2_mean])
# Step 3. Attentive-Matching
# Each forward (or backward) similarity is taken as the weight
# of the forward (or backward) contextual embedding, and calculate an
# attentive vector for the sentence by weighted summing all its
# contextual embeddings.
# Finally match each forward (or backward) contextual embedding
# with its corresponding attentive vector.
# (batch, seq_len1, seq_len2, hidden_dim)
att_2 = context_2.unsqueeze(-3) * cosine_sim.unsqueeze(-1)
# (batch, seq_len1, seq_len2, hidden_dim)
att_1 = context_1.unsqueeze(-2) * cosine_sim.unsqueeze(-1)
if self.with_attentive_match:
# (batch, seq_len*, hidden_dim)
att_mean_2 = masked_softmax(att_2.sum(dim=2), mask_1.unsqueeze(-1))
att_mean_1 = masked_softmax(att_1.sum(dim=1), mask_2.unsqueeze(-1))
# (batch, seq_len*, num_perspectives)
matching_vector_1_att_mean = multi_perspective_match(context_1,
att_mean_2,
self.attentive_match_weights)
matching_vector_2_att_mean = multi_perspective_match(context_2,
att_mean_1,
self.attentive_match_weights_reversed)
matching_vector_1.extend(matching_vector_1_att_mean)
matching_vector_2.extend(matching_vector_2_att_mean)
# Step 4. Max-Attentive-Matching
# Pick the contextual embeddings with the highest cosine similarity as the attentive
# vector, and match each forward (or backward) contextual embedding with its
# corresponding attentive vector.
if self.with_max_attentive_match:
# (batch, seq_len*, hidden_dim)
att_max_2 = masked_max(att_2, mask_2.unsqueeze(-2).unsqueeze(-1), dim=2)
att_max_1 = masked_max(att_1.permute(0, 2, 1, 3), mask_1.unsqueeze(-2).unsqueeze(-1), dim=2)
# (batch, seq_len*, num_perspectives)
matching_vector_1_att_max = multi_perspective_match(context_1,
att_max_2,
self.max_attentive_match_weights)
matching_vector_2_att_max = multi_perspective_match(context_2,
att_max_1,
self.max_attentive_match_weights_reversed)
matching_vector_1.extend(matching_vector_1_att_max)
matching_vector_2.extend(matching_vector_2_att_max)
return matching_vector_1, matching_vector_2
def sort_and_run_forward(self,
module: Callable[[PackedSequence, Optional[RnnState]],
Tuple[Union[PackedSequence, torch.Tensor], RnnState]],
inputs: torch.Tensor,
mask: torch.Tensor,
hidden_state: Optional[RnnState] = None):
"""
This function exists because Pytorch RNNs require that their inputs be sorted
before being passed as input. As all of our Seq2xxxEncoders use this functionality,
it is provided in a base class. This method can be called on any module which
takes as input a ``PackedSequence`` and some ``hidden_state``, which can either be a
tuple of tensors or a tensor.
As all of our Seq2xxxEncoders have different return types, we return `sorted`
outputs from the module, which is called directly. Additionally, we return the
indices into the batch dimension required to restore the tensor to it's correct,
unsorted order and the number of valid batch elements (i.e the number of elements
in the batch which are not completely masked). This un-sorting and re-padding
of the module outputs is left to the subclasses because their outputs have different
types and handling them smoothly here is difficult.
Parameters
----------
module : ``Callable[[PackedSequence, Optional[RnnState]],
Tuple[Union[PackedSequence, torch.Tensor], RnnState]]``, required.
A function to run on the inputs. In most cases, this is a ``torch.nn.Module``.
inputs : ``torch.Tensor``, required.
A tensor of shape ``(batch_size, sequence_length, embedding_size)`` representing
the inputs to the Encoder.
mask : ``torch.Tensor``, required.
A tensor of shape ``(batch_size, sequence_length)``, representing masked and
non-masked elements of the sequence for each element in the batch.
hidden_state : ``Optional[RnnState]``, (default = None).
A single tensor of shape (num_layers, batch_size, hidden_size) representing the
state of an RNN with or a tuple of
tensors of shapes (num_layers, batch_size, hidden_size) and
(num_layers, batch_size, memory_size), representing the hidden state and memory
state of an LSTM-like RNN.
Returns
-------
module_output : ``Union[torch.Tensor, PackedSequence]``.
A Tensor or PackedSequence representing the output of the Pytorch Module.
The batch size dimension will be equal to ``num_valid``, as sequences of zero
length are clipped off before the module is called, as Pytorch cannot handle
zero length sequences.
final_states : ``Optional[RnnState]``
A Tensor representing the hidden state of the Pytorch Module. This can either
be a single tensor of shape (num_layers, num_valid, hidden_size), for instance in
the case of a GRU, or a tuple of tensors, such as those required for an LSTM.
restoration_indices : ``torch.LongTensor``
A tensor of shape ``(batch_size,)``, describing the re-indexing required to transform
the outputs back to their original batch order.
"""
# In some circumstances you may have sequences of zero length. ``pack_padded_sequence``
# requires all sequence lengths to be > 0, so remove sequences of zero length before
# calling self._module, then fill with zeros.
# First count how many sequences are empty.
batch_size = mask.size(0)
num_valid = torch.sum(mask[:, 0]).int().item()
sequence_lengths = get_lengths_from_binary_sequence_mask(mask)
sorted_inputs, sorted_sequence_lengths, restoration_indices, sorting_indices =\
sort_batch_by_length(inputs, sequence_lengths)
# Now create a PackedSequence with only the non-empty, sorted sequences.
packed_sequence_input = pack_padded_sequence(sorted_inputs[:num_valid, :, :],
sorted_sequence_lengths[:num_valid].data.tolist(),
batch_first=True)
# Prepare the initial states.
if not self.stateful:
if hidden_state is None:
initial_states = hidden_state
elif isinstance(hidden_state, tuple):
initial_states = [state.index_select(1, sorting_indices)[:, :num_valid, :].contiguous()
for state in hidden_state]
else:
initial_states = hidden_state.index_select(1, sorting_indices)[:, :num_valid, :].contiguous()
else:
initial_states = self._get_initial_states(batch_size, num_valid, sorting_indices)
# Actually call the module on the sorted PackedSequence.
module_output, final_states = module(packed_sequence_input, initial_states)
return module_output, final_states, restoration_indices
def forward(
self, # type: ignore
tokens: Dict[str, torch.LongTensor],
spans: torch.LongTensor,
metadata: List[Dict[str, Any]],
pos_tags: Dict[str, torch.LongTensor] = None,
span_labels: torch.LongTensor = None) -> Dict[str, torch.Tensor]:
# pylint: disable=arguments-differ
"""
Parameters
----------
tokens : Dict[str, torch.LongTensor], required
The output of ``TextField.as_array()``, which should typically be passed directly to a
``TextFieldEmbedder``. This output is a dictionary mapping keys to ``TokenIndexer``
tensors. At its most basic, using a ``SingleIdTokenIndexer`` this is: ``{"tokens":
Tensor(batch_size, num_tokens)}``. This dictionary will have the same keys as were used
for the ``TokenIndexers`` when you created the ``TextField`` representing your
sequence. The dictionary is designed to be passed directly to a ``TextFieldEmbedder``,
which knows how to combine different word representations into a single vector per
token in your input.
spans : ``torch.LongTensor``, required.
A tensor of shape ``(batch_size, num_spans, 2)`` representing the
inclusive start and end indices of all possible spans in the sentence.
metadata : List[Dict[str, Any]], required.
A dictionary of metadata for each batch element which has keys:
tokens : ``List[str]``, required.
The original string tokens in the sentence.
gold_tree : ``nltk.Tree``, optional (default = None)
Gold NLTK trees for use in evaluation.
pos_tags : ``List[str]``, optional.
The POS tags for the sentence. These can be used in the
model as embedded features, but they are passed here
in addition for use in constructing the tree.
pos_tags : ``torch.LongTensor``, optional (default = None)
The output of a ``SequenceLabelField`` containing POS tags.
span_labels : ``torch.LongTensor``, optional (default = None)
A torch tensor representing the integer gold class labels for all possible
spans, of shape ``(batch_size, num_spans)``.
Returns
-------
An output dictionary consisting of:
class_probabilities : ``torch.FloatTensor``
A tensor of shape ``(batch_size, num_spans, span_label_vocab_size)``
representing a distribution over the label classes per span.
spans : ``torch.LongTensor``
The original spans tensor.
tokens : ``List[List[str]]``, required.
A list of tokens in the sentence for each element in the batch.
pos_tags : ``List[List[str]]``, required.
A list of POS tags in the sentence for each element in the batch.
num_spans : ``torch.LongTensor``, required.
A tensor of shape (batch_size), representing the lengths of non-padded spans
in ``enumerated_spans``.
loss : ``torch.FloatTensor``, optional
A scalar loss to be optimised.
"""
embedded_text_input = self.text_field_embedder(tokens)
if pos_tags is not None and self.pos_tag_embedding is not None:
embedded_pos_tags = self.pos_tag_embedding(pos_tags)
embedded_text_input = torch.cat(
[embedded_text_input, embedded_pos_tags], -1)
elif self.pos_tag_embedding is not None:
raise ConfigurationError(
"Model uses a POS embedding, but no POS tags were passed.")
mask = get_text_field_mask(tokens)
# Looking at the span start index is enough to know if
# this is padding or not. Shape: (batch_size, num_spans)
span_mask = (spans[:, :, 0] >= 0).squeeze(-1).long()
if span_mask.dim() == 1:
# This happens if you use batch_size 1 and encounter
# a length 1 sentence in PTB, which do exist. -.-
span_mask = span_mask.unsqueeze(-1)
if span_labels is not None and span_labels.dim() == 1:
span_labels = span_labels.unsqueeze(-1)
num_spans = get_lengths_from_binary_sequence_mask(span_mask)
encoded_text = self.encoder(embedded_text_input, mask)
span_representations = self.span_extractor(encoded_text, spans, mask,
span_mask)
if self.feedforward_layer is not None:
span_representations = self.feedforward_layer(span_representations)
logits = self.tag_projection_layer(span_representations)
class_probabilities = last_dim_softmax(logits, span_mask.unsqueeze(-1))
output_dict = {
"class_probabilities": class_probabilities,
"spans": spans,
"tokens": [meta["tokens"] for meta in metadata],
"pos_tags": [meta.get("pos_tags") for meta in metadata],
"num_spans": num_spans
}
if span_labels is not None:
loss = sequence_cross_entropy_with_logits(logits, span_labels,
span_mask)
self.tag_accuracy(class_probabilities, span_labels, span_mask)
output_dict["loss"] = loss
# The evalb score is expensive to compute, so we only compute
# it for the validation and test sets.
batch_gold_trees = [meta.get("gold_tree") for meta in metadata]
if all(batch_gold_trees
) and self._evalb_score is not None and not self.training:
gold_pos_tags: List[List[str]] = [
list(zip(*tree.pos()))[1] for tree in batch_gold_trees
]
predicted_trees = self.construct_trees(
class_probabilities.cpu().data,
spans.cpu().data, num_spans.data, output_dict["tokens"],
gold_pos_tags)
self._evalb_score(predicted_trees, batch_gold_trees)
return output_dict
def __call__(self,
predictions: torch.Tensor,
gold_labels: torch.Tensor,
mask: Optional[torch.Tensor] = None,
prediction_map: Optional[torch.Tensor] = None):
"""
Parameters
----------
predictions : ``torch.Tensor``, required.
A tensor of predictions of shape (batch_size, sequence_length, num_classes).
gold_labels : ``torch.Tensor``, required.
A tensor of integer class label of shape (batch_size, sequence_length). It must be the same
shape as the ``predictions`` tensor without the ``num_classes`` dimension.
mask: ``torch.Tensor``, optional (default = None).
A masking tensor the same size as ``gold_labels``.
prediction_map: ``torch.Tensor``, optional (default = None).
A tensor of size (batch_size, num_classes) which provides a mapping from the index of predictions
to the indices of the label vocabulary. If provided, the output label at each timestep will be
``vocabulary.get_index_to_token_vocabulary(prediction_map[batch, argmax(predictions[batch, t]))``,
rather than simply ``vocabulary.get_index_to_token_vocabulary(argmax(predictions[batch, t]))``.
This is useful in cases where each Instance in the dataset is associated with a different possible
subset of labels from a large label-space (IE FrameNet, where each frame has a different set of
possible roles associated with it).
"""
if mask is None:
mask = torch.ones_like(gold_labels)
predictions, gold_labels, mask, prediction_map = self.unwrap_to_tensors(predictions,
gold_labels,
mask, prediction_map)
num_classes = predictions.size(-1)
if (gold_labels >= num_classes).any():
raise ConfigurationError("A gold label passed to SpanBasedF1Measure contains an "
"id >= {}, the number of classes.".format(num_classes))
sequence_lengths = get_lengths_from_binary_sequence_mask(mask)
argmax_predictions = predictions.max(-1)[1]
if prediction_map is not None:
argmax_predictions = torch.gather(prediction_map, 1, argmax_predictions)
gold_labels = torch.gather(prediction_map, 1, gold_labels.long())
argmax_predictions = argmax_predictions.float()
# Iterate over timesteps in batch.
batch_size = gold_labels.size(0)
for i in range(batch_size):
sequence_prediction = argmax_predictions[i, :]
sequence_gold_label = gold_labels[i, :]
length = sequence_lengths[i]
if length == 0:
# It is possible to call this metric with sequences which are
# completely padded. These contribute nothing, so we skip these rows.
continue
predicted_string_labels = [self._label_vocabulary[label_id]
for label_id in sequence_prediction[:length].tolist()]
gold_string_labels = [self._label_vocabulary[label_id]
for label_id in sequence_gold_label[:length].tolist()]
if self._label_encoding == "BIO":
predicted_spans = bio_tags_to_spans(predicted_string_labels, self._ignore_classes)
gold_spans = bio_tags_to_spans(gold_string_labels, self._ignore_classes)
elif self._label_encoding == "IOB1":
predicted_spans = iob1_tags_to_spans(predicted_string_labels, self._ignore_classes)
gold_spans = iob1_tags_to_spans(gold_string_labels, self._ignore_classes)
elif self._label_encoding == "BIOUL":
predicted_spans = bioul_tags_to_spans(predicted_string_labels, self._ignore_classes)
gold_spans = bioul_tags_to_spans(gold_string_labels, self._ignore_classes)
predicted_spans = self._handle_continued_spans(predicted_spans)
gold_spans = self._handle_continued_spans(gold_spans)
for span in predicted_spans:
if span in gold_spans:
self._true_positives[span[0]] += 1
gold_spans.remove(span)
else:
self._false_positives[span[0]] += 1
# These spans weren't predicted.
for span in gold_spans:
self._false_negatives[span[0]] += 1
