Python LongTensor.float方法代码示例

# 需要导入模块: from torch import LongTensor [as 别名]
# 或者: from torch.LongTensor import float [as 别名]
    def _prepare_decode_step_input(self,
                                   input_indices: torch.LongTensor,
                                   decoder_hidden_state: torch.LongTensor = None,
                                   encoder_outputs: torch.LongTensor = None,
                                   encoder_outputs_mask: torch.LongTensor = None) -> torch.LongTensor:
        """
        Given the input indices for the current timestep of the decoder, and all the encoder
        outputs, compute the input at the current timestep.  Note: This method is agnostic to
        whether the indices are gold indices or the predictions made by the decoder at the last
        timestep. So, this can be used even if we're doing some kind of scheduled sampling.

        If we're not using attention, the output of this method is just an embedding of the input
        indices.  If we are, the output will be a concatentation of the embedding and an attended
        average of the encoder inputs.

        Parameters
        ----------
        input_indices : torch.LongTensor
            Indices of either the gold inputs to the decoder or the predicted labels from the
            previous timestep.
        decoder_hidden_state : torch.LongTensor, optional (not needed if no attention)
            Output of from the decoder at the last time step. Needed only if using attention.
        encoder_outputs : torch.LongTensor, optional (not needed if no attention)
            Encoder outputs from all time steps. Needed only if using attention.
        encoder_outputs_mask : torch.LongTensor, optional (not needed if no attention)
            Masks on encoder outputs. Needed only if using attention.
        """
        # input_indices : (batch_size,)  since we are processing these one timestep at a time.
        # (batch_size, target_embedding_dim)
        embedded_input = self._target_embedder(input_indices)
        if self._attention_function:
            # encoder_outputs : (batch_size, input_sequence_length, encoder_output_dim)
            # Ensuring mask is also a FloatTensor. Or else the multiplication within attention will
            # complain.
            encoder_outputs_mask = encoder_outputs_mask.float()
            # (batch_size, input_sequence_length)
            input_weights = self._decoder_attention(decoder_hidden_state, encoder_outputs, encoder_outputs_mask)
            # (batch_size, encoder_output_dim)
            attended_input = weighted_sum(encoder_outputs, input_weights)
            # (batch_size, encoder_output_dim + target_embedding_dim)
            return torch.cat((attended_input, embedded_input), -1)
        else:
            return embedded_input

# 需要导入模块: from torch import LongTensor [as 别名]
# 或者: from torch.LongTensor import float [as 别名]
    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 = util.ones_like(sequence_tensor[:, 0, 0]).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)
#.........这里部分代码省略.........

# 需要导入模块: from torch import LongTensor [as 别名]
# 或者: from torch.LongTensor import float [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