How to do large-scale outer product efficiently

[LilySnow (Li Xue)]

I have two matrices A and B. A and B have the same number of rows (m), and different number of columns. A.shape = [m,M], B.shape = [m,N]. For each row of A and B, I want to do an outer product (torch.outer).

Is there a way not using loop, because m is very large and I need to do this operation many times.

Thanks.

[ZimoNitrome]

https://pytorch.org/docs/stable/generated/torch.outer.html

[LilySnow (Li Xue)]

torch.outer only works on two vectors

[ZimoNitrome]

This seems to work if you want to just do it for 2D tensors.

Batch Outer Product

No, there is no batch version for ger. But you can obtain the same results by appending 1 to the tensors. x = torch.rand(10, 3) y = torch.rand(10, 3) res = torch.bmm(x.unsqueeze(2), y.unsqueeze(1))

[AlphaBetaGamma96]

if you have a tensor A of shape [m,M] and another tensor B of shape [m,N] you can do an outer product via torch.einsum,

outer_product = torch.einsum("bi,bj->bij",A,B)