Product of two tensors of size 4

1

Hi,
u= torch.randn(3, 1, 40, 64)
v= torch.randn(1, 3, 40, 64)
hawo can I compute the product of u and v.torch.bmm not work for this example.

Thanks,

2

What do you mean by product here? matrix-matrix multiplication? Or broadcasting + element-wise?

3

@albanD
Multiplication

4

What is the output size you expect?? [3, 40, 64]? [1, 40 64]? [3, 3, 40]?

5

Let me guess…
Do you mean

u -> (A, B, batch_1, batch_2)
b -> (B, A, batch_1, batch_2)
result -> (A, A, batch_1, batch_2)

which result[:, :, i, j] equals to u[:, :, i, j] @ v[:, :, i, j]?
Just like a 4-dim bmm?

6

Hi@Eta_C
Exactly the result will be :
(A, A, batch_1, batch_2)
What do you mean by i and j?

7

If I understand correctly…
Try this, you should build a small example to test its correctness.

import torch

u = torch.randn(3, 1, 40, 64)
v = torch.randn(1, 3, 40, 64)
_u = u.reshape(3, 1, -1)
_v = v.reshape(1, 3, -1)
_result = torch.bmm(_u.permute(2, 0, 1), _v.permute(2, 0, 1))
result = _result.permute(1, 2, 0).reshape(3, 3, 40, 64)
1 Like