Python Tensor.gather方法代码示例

# 需要导入模块: from torch import Tensor [as 别名]
# 或者: from torch.Tensor import gather [as 别名]
def get_final_encoder_states(encoder_outputs: torch.Tensor,
                             mask: torch.Tensor,
                             bidirectional: bool = False) -> torch.Tensor:
    """
    Given the output from a ``Seq2SeqEncoder``, with shape ``(batch_size, sequence_length,
    encoding_dim)``, this method returns the final hidden state for each element of the batch,
    giving a tensor of shape ``(batch_size, encoding_dim)``.  This is not as simple as
    ``encoder_outputs[:, -1]``, because the sequences could have different lengths.  We use the
    mask (which has shape ``(batch_size, sequence_length)``) to find the final state for each batch
    instance.

    Additionally, if ``bidirectional`` is ``True``, we will split the final dimension of the
    ``encoder_outputs`` into two and assume that the first half is for the forward direction of the
    encoder and the second half is for the backward direction.  We will concatenate the last state
    for each encoder dimension, giving ``encoder_outputs[:, -1, :encoding_dim/2]`` concated with
    ``encoder_outputs[:, 0, encoding_dim/2:]``.
    """
    # These are the indices of the last words in the sequences (i.e. length sans padding - 1).  We
    # are assuming sequences are right padded.
    # Shape: (batch_size,)
    last_word_indices = mask.sum(1).long() - 1
    batch_size, _, encoder_output_dim = encoder_outputs.size()
    expanded_indices = last_word_indices.view(-1, 1, 1).expand(batch_size, 1, encoder_output_dim)
    # Shape: (batch_size, 1, encoder_output_dim)
    final_encoder_output = encoder_outputs.gather(1, expanded_indices)
    final_encoder_output = final_encoder_output.squeeze(1)  # (batch_size, encoder_output_dim)
    if bidirectional:
        final_forward_output = final_encoder_output[:, :(encoder_output_dim // 2)]
        final_backward_output = encoder_outputs[:, 0, (encoder_output_dim // 2):]
        final_encoder_output = torch.cat([final_forward_output, final_backward_output], dim=-1)
    return final_encoder_output

# 需要导入模块: from torch import Tensor [as 别名]
# 或者: from torch.Tensor import gather [as 别名]
    def _joint_likelihood(self,
                          logits: torch.Tensor,
                          tags: torch.Tensor,
                          mask: torch.LongTensor) -> torch.Tensor:
        """
        Computes the numerator term for the log-likelihood, which is just score(inputs, tags)
        """
        batch_size, sequence_length, num_tags = logits.data.shape

        # Transpose batch size and sequence dimensions:
        logits = logits.transpose(0, 1).contiguous()
        mask = mask.float().transpose(0, 1).contiguous()
        tags = tags.transpose(0, 1).contiguous()

        # Start with the transition scores from start_tag to the first tag in each input
        if self.include_start_end_transitions:
            score = self.start_transitions.index_select(0, tags[0])
        else:
            score = 0.0

        # Broadcast the transition scores to one per batch element
        broadcast_transitions = self.transitions.view(1, num_tags, num_tags).expand(batch_size, num_tags, num_tags)

        # Add up the scores for the observed transitions and all the inputs but the last
        for i in range(sequence_length - 1):
            # Each is shape (batch_size,)
            current_tag, next_tag = tags[i], tags[i+1]

            # The scores for transitioning from current_tag to next_tag
            transition_score = (
                    broadcast_transitions
                    # Choose the current_tag-th row for each input
                    .gather(1, current_tag.view(batch_size, 1, 1).expand(batch_size, 1, num_tags))
                    # Squeeze down to (batch_size, num_tags)
                    .squeeze(1)
                    # Then choose the next_tag-th column for each of those
                    .gather(1, next_tag.view(batch_size, 1))
                    # And squeeze down to (batch_size,)
                    .squeeze(1)
            )

            # The score for using current_tag
            emit_score = logits[i].gather(1, current_tag.view(batch_size, 1)).squeeze(1)

            # Include transition score if next element is unmasked,
            # input_score if this element is unmasked.
            score = score + transition_score * mask[i + 1] + emit_score * mask[i]

        # Transition from last state to "stop" state. To start with, we need to find the last tag
        # for each instance.
        last_tag_index = mask.sum(0).long() - 1
        last_tags = tags.gather(0, last_tag_index.view(1, batch_size).expand(sequence_length, batch_size))

        # Is (sequence_length, batch_size), but all the columns are the same, so take the first.
        last_tags = last_tags[0]

        # Compute score of transitioning to `stop_tag` from each "last tag".
        if self.include_start_end_transitions:
            last_transition_score = self.end_transitions.index_select(0, last_tags)
        else:
            last_transition_score = 0.0

        # Add the last input if it's not masked.
        last_inputs = logits[-1]                                         # (batch_size, num_tags)
        last_input_score = last_inputs.gather(1, last_tags.view(-1, 1))  # (batch_size, 1)
        last_input_score = last_input_score.squeeze()                    # (batch_size,)

        score = score + last_transition_score + last_input_score * mask[-1]

        return score

# 需要导入模块: from torch import Tensor [as 别名]
# 或者: from torch.Tensor import gather [as 别名]
    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.
#.........这里部分代码省略.........