You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I think the next big step after #17 is adding multiset-valued tensors.
Steps:
Add multisets: these are simply collections where multiplicity counts but not order. I don't think we'll need to add a comparison operator to them, so I think we could just implement them using arrays.
Add the bar function (or whatever you wanna call it, in the paper it's represented as a bar above its operand), which is a function taking a tensor from A to B and returning a tensor from multisets(A) to multisets(B).
Make reduction functions reduce multisets into single values. E.g: sum({{1, 3, 3}}) = 7 (where {{1, 3, 3}} is the multiset with 1, 3 and 3).
With those two things, reductions with index mapping f, tensor U and reduction fn omega, can be implemented as: omega ○ bar(U) ○ f
Activity