Skip to content

RETURNN common principles

Jump to bottom
Albert Zeyer edited this page Apr 29, 2022 · 3 revisions

It inherits from the RETURNN principles.

But further:

  • The fact that the order of dimensions does not matter is emphasized even much more. Tensor.shape is a set of dims. The order really should never matter. You can never access an axis by integer.

  • For any function: Any dims which are not referred to are treated as batch dims. There could be multiple batch-like dims.

Clone this wiki locally