[FR] Optional channels

[ehknight]

Sometimes, I find myself working lists of tensors in which one tensor has a shape (b, c) (for c classes) and another tensor has shape (b,) (for a single class). My current approach is to pad the tensors that have only one class with an additional channel dimension, use rearrange on the list, and then squeeze the dimensions that need to be squeezed.

A great alternative to this would be supporting optional channels. Perhaps you could notate them with a question mark: rearrange(x, "b c? -> (b c?)").

[arogozhnikov]

@ehknight thanks for coming up with suggestion.

I've been previously considering this feature, however it is quite difficult to generalize beyond a single optional dimension.
The way your example should be implemented in existing einops syntax is:

rearrange(x, "b ... -> (b ...)")

where ellipsis can be 0 or 1 dimensions (or more). However, right now that's not yet supported and is subject of #43

[arogozhnikov]

@ehknight that's implemented now and will become a part of the next release

[shoyer]

It looks like transformations like "b ... -> (b ...)" work currently, but not "(b ...) -> b ..."?

[arogozhnikov]

@shoyer that's right, those expressions are disallowed.

Let's analyze your example:

(b ...) 

left part implies there is one dimension which should be decomposed in b and undefined number of dimensions.
To make this a valid operation one needs to provide number of dimensions in ... and their sizes.
In this scenario user can write ... as separate axes like (b b1 b2) -> b b1 b2.

[arogozhnikov]

Ellipsis collapsing is now available as a part of einops 0.3 release.