c84c714c64f85c01c2b748fd035a19e6e9154648,allennlp/modules/seq2seq_encoders/multi_head_self_attention.py,MultiHeadSelfAttention,forward,#MultiHeadSelfAttention#Any#Any#,78

Before Change

        // Note that we _cannot_ use a reshape here, because this tensor was created
        // with num_heads being the first dimension, so reshaping naively would not
        // throw an error, but give an incorrect result.
        outputs = torch.cat(torch.split(outputs, batch_size, dim=0), dim=-1)

        // Project back to original input size.
        // shape (batch_size, timesteps, input_size)
        outputs = self._output_projection(outputs)

After Change

        // shape (batch_size, num_heads, timesteps, values_dim/num_heads)
        outputs = outputs.view(batch_size, num_heads, timesteps, int(self._values_dim / num_heads))
        // shape (batch_size, timesteps, num_heads, values_dim/num_heads)
        outputs = outputs.transpose(1, 2).contiguous()
        // shape (batch_size, timesteps, values_dim)
        outputs = outputs.view(batch_size, timesteps, self._values_dim)

        // Project back to original input size.

Instances