With two tensors
a = torch.ones([256, 512, 32])
b = torch.ones([32, 2])
what is the most efficient way to broadcast b onto every associated entry in a, producing a result with shape
[256, 512, 32, 2] ?
Is there an inplace variant maybe? I want to avoid making copies as much as possible…
a.unsqueeze(-1).mul(b.unsqueeze(0).unsqueeze(0))
But it’s sort of hard to see in your example because you are multiplying everything by ones. You can open a REPL. I also think with pytorch 1.3 you could use named tensors and just make sure that a's dimensions are named something like N, C, H and b's dimensions are H, W… basically just ensure that the dimensions align correctly.
thank you, but do I really need those last two unsqueezes?
a = torch.ones([256, 512, 32]).unsqueeze(-1)
b = torch.ones([32, 2])
a * b
also, wouldnt be mul_ be more efficient here?
The output is not the same size as either of the inputs. So you can’t really do it inplace.