einsum notation

[JohannesBuchner]

Hi @dkogan,

Thanks for this very interesting project, I came across it from HN and your talk.

1. I was looking at the syntax for broadcast_define, which is given by
   @nps.broadcast_define( (('n',), ('n',)) ) or @nps.broadcast_define( (('n',), ('n',), ())) in the examples.

I am not sure whether you are familiar with numpy's einsum function. I quite like the intuitive style of writing indices.

I am wondering whether it could be reused here. In that case, one could write something like:
@nps.broadcast_define('n,n->')
@nps.broadcast_define('n,mn->m')

However, I am not sure how complex the parser of einsum is and whether it can be easily reused here.

2. Maybe one could also consider a concat/stacking API that is based on einsum notation:
   cat(arr13,arr3) -> einstack(arr13,arr3,'ij,j->kij')
glue(arr13,arr3) -> einstack(arr13,arr3,'ij,ij->ij')

3. I am a bit scared of the automatic adding of leading dimensions, because it may hide passing the wrong array. Is there a way to turn it off?

[dkogan]

Hello!

I am not sure whether you are familiar with numpy's einsum function. I quite like the intuitive style of writing indices. I am wondering whether it could be reused here. In that case, one could write something like: @nps.broadcast_define('n,n->') @nps.broadcast_define('n,mn->m') 2 Maybe one could also consider a concat/stacking API that is based on einsum notation: cat(arr13,arr3) -> einstack(arr13,arr3,'ij,j->kij') glue(arr13,arr3) -> einstack(arr13,arr3,'ij,ij->ij')

Hi. I don't have plans for any new features for numpysane right now (busy with other things). But if you want to contribute some of this, we can certainly talk about it.

3 I am a bit scared of the automatic adding of leading dimensions, because it may hide passing the wrong array. Is there a way to turn it off?

No. This is a core part of the broadcasting logic. This does imply one particular interpretation of the input, as you say, but it enables MUCH nicer code, which makes it worth it, in my opinion.