RuntimeError: Expected 3-dimensional tensor, but got 2-dimensional tensor for argument

Question

I have two tensors names: 'wy' and 'x', both of them with size 8:

wy= tensor([[ 74.2090, -92.9444,  45.2677, -38.4132, -39.8641,  -6.9193,  67.4830,
     -80.1534]], grad_fn=<AddmmBackward>)

x= tensor([[ 70.,  77., 101.,  75.,  40.,  83.,  48.,  73.]])

Now, I want to do bmm to multiply x * wy as follow:

 xWy = x.bmm(Wy.unsqueeze(2)).squeeze(3)

I got an error:

  RuntimeError: Expected 3-dimensional tensor, but got 2-dimensional tensor for argument #1 'batch1' (while checking arguments for bmm)

8*8 should be possible. but I don't know why I got this error every time.

any help, please!

So you want `(1, 8) * (8, 1)` that results to a scalar? – user2246849 May 19 '22 at 05:59 — user2246849, May 19 '22 at 05:59
Alright, see my answer below then! – user2246849 May 19 '22 at 09:43 — user2246849, May 19 '22 at 09:43
This has nothing to do with AllenNLP. – Dirk Groeneveld May 26 '22 at 20:23 — Dirk Groeneveld, May 26 '22 at 20:23

user2246849 · Accepted Answer · 2022-05-19T11:08:32.997

bmm stands for batch matrix-matrix product. So it expects both tensors with a batch dimension (i.e., 3D as the error says).

For single tensors, you want to use mm instead. Note that Tensor.mm() also exists with the same behaviour.

x.mm(wy.transpose(0, 1))

tensor([[-5051.9199]])

Or better, for two 1D tensor you can use dot for dot product.

# Or simply do not initialise them with an additional dimension. Not needed.
x.squeeze().dot(wy.squeeze())

tensor(-5051.9199)

RuntimeError: Expected 3-dimensional tensor, but got 2-dimensional tensor for argument

1 Answers1