Is there a design choice for why cubes can’t be directly handled by a transform layer? They are, after all, already Tensor objects and the semantics seem unambiguous to me. Is it a “no time to do it just yet” issue? Or is there some deeper reason that I’m missing?

The question I’m asking is answered here if anyone needs the “answer”.

it’s sort of ambiguous on what you want to do with a cube. Would one do batch matrix multiply (like torch.bmm) or would one do broadcasted matrix multiply…

Either ways, it hasn’t been thought through and implemented.

Interesting, I hadn’t thought of the ambiguity. I guess I assumed that batch matrix multiply was the natural extension of 2d matrix multiplication, but I can see your point.

Thanks for the reply. And I understand this isn’t an issue since the operation can already be achieved.