Skip to content

scaled_dot_product_matrix_attention

allennlp.modules.matrix_attention.scaled_dot_product_matrix_attention

[SOURCE]


ScaledDotProductMatrixAttention

@MatrixAttention.register("scaled_dot_product")
class ScaledDotProductMatrixAttention(DotProductMatrixAttention)

Computes attention between every entry in matrix_1 with every entry in matrix_2 using a dot product. Scales the result by the size of the embeddings.

Registered as a MatrixAttention with name "scaled_dot_product".

forward

class ScaledDotProductMatrixAttention(DotProductMatrixAttention):
 | ...
 | def forward(
 |     self,
 |     matrix_1: torch.Tensor,
 |     matrix_2: torch.Tensor
 | ) -> torch.Tensor