Mike J Innes
Mike J Innes
I'm curious why/if `Buffer` doesn't do the job here. It seems like it should solve the main problem – huge number of array allocations – quite neatly, even if there's...
Yeah, that's likely to be the issue if the work inside the body of the loop is small. Type inference in the presence of branches is effectively disabled pending compiler...
Awesome, really nice work @ToucheSir. If this is based on @Keno's original code it probably makes sense to add a co-author to the commit? (Alternatively you could treat this as...
In general it'd be cleaner if we could avoid defining second order gradients in favour of e.g. having dgetindex use `setindex` (out of place) and adding a first-order adjoint for...
Sounds fine to me. This is kind of a tricky tradeoff unfortunately; probably the only _real_ answer is to delete the adjoint entirely and support differentiating through map, but that...
I have no issue with this if it works. If no one objects I'll merge in a few days.
The goal is a bit different from `StaticArray`s since we don't expect them to be specialised on size; they should behave exactly like normal `Array`s that happen to be immutable....
This seems like a good idea, but I think we could do a slightly different setup: * Have a global var for determinism, and provide getters and setters for this...
Looks fine to me, thanks. Would like a thumbs up from @maleadt.
As I understand it, `ho` should already have been expanded (via `hbatch`) during the forwards pass. Its gradient `dho` should have the same size as `ho`, in which case AFAICT...