Few implementations that do not match the original paper.

[siaoliu]

The similarity computation is cosine-related function in the paper, but here is torch.bmm, may be such function is not important.

[etasnadi]

It is the cosine similarity in the paper that is usually computed using dot product, therefore bmm may compute the cos-similarity of each possible vector pairs ([cosine] similarity matrix) coming from the output of the backbone net I guess.