Broadcast Tensor to Have Certain Size

1

Hi!

I have a tensor variable x, that may come in sizes (,), (n,), (1, n) or (m,n).
Independently of the particular case, I would like to explicitly broadcast x to be two-dimensional.
In other words, is there a single call like y = x.broadcast_to_dim(2) that will do nothing if x has size (1,n) or (m,n), and will insert an axis on the left if x has size (n,) and two axes on the left if x has size (,), making x of size (1,n) in both cases?

Thanks a lot!

2

Does something like

if x.dim() == 1:
    x = x.epxand(1, n)

work?

Best regards

Tomas

3

Thanks!

I now did

if x.dim() <= 1:
    x = x.expand(1, *(x.size() or [1]))

which does the job, but unfortunately must be part of my forward method.
What kind of toll does this check have on performance?

4

Not much, as it is metadata only. But you might try expand(1, -1) instead of the complicated thing or treat dim() == 0 and 1 separately.