Pytorch: tensor multiplication

1

Hi!
I want to code this matrix multiplication:

image
B=16, h=32, w=56, b*b=64, n=25
my code:

torch.mul(torch.rand([16, 32, 56, 64]),torch.rand([16, 32, 56, 64,25]))

but I encountered in with this error:

The size of tensor a (64) must match the size of tensor b (25) at non-singleton dimension 4

Would you mind helping me to fix this problem?

2

Hi,

I think the simplest is to write it as a matrix matrix multiplication. And so add a temporary dimension of size 1 on the first Tensor:

torch.matmul(torch.rand([16, 32, 56, 1, 64]),torch.rand([16, 32, 56, 64,25])).squeeze(-2)
1 Like
3

Thanks for your reply. The problem solved.
Would you mind explaining it? I do not understand it.

4

It consider Bxhxw as batch dimensions. And 1x(bb) and (bb)xn for the other dimensions of each other dimensions.
Doing the matrix multiply between the two, will give you a Tensor with 1xn and with the same batch size.
The squeeze removes the dimension of size 1.

1 Like
5

Thanks for your help dear @albanD :slight_smile: