Michiel Stock
Michiel Stock
Strange, can you give a small working example?
Maybe you can best take the lead in this and we can look together if it can be improved. Feel free to open a PR!
Still stuck on this, why does computing gradients work for `logdet` but not `tr` or `sum`. It should just fall back to the simple shortcuts, for which adjoints already exist?
Technically, it only makes sense to define the adjoints for those function where Kronecker provides shortcuts, based on this rule: https://en.wikipedia.org/wiki/Matrix_calculus#Identities_in_differential_form
I will look at it Friday during the hackathon!
Can we discuss irl somewhere this week? Have some questions. Btw, `kron` from Zygote seems to work a bit faster than base.
I don't see the value of generating and storing two dense matrices? If we collect as you say, we better generate them on the fly once? Especially for in-place collecting,...
> I agree that storing the identity matrices as dense arrays is highly inefficient. Although I don't show it in my example, my idea was that if `A` or `B`...
Yes, it cannot access the fast multiplication this way. I have added a naive decomposition for this, so your examples is now: ```julia julia> @time kronecker(A+A,C)*ones(20000); 0.011665 seconds (41.67 k...
I understand the problem, but the very idea of `Kronecker` is that it exports `⊗` as a new matrix product so I would like it to export this by default....