fill can't handle nothing as a recieved gradient

[dsweber2]

I was using fill as a way of indexing into another matrix, e.g. x[fill(1,N)...], and this causes (nothing,) to be handed to the pullback of fill. As all the gradient does is sum at the moment, this causes errors trying to sum nothing. I address this by checking the type, and passing on any received nothing, and if the entries are mixed it ignores any nothings in the sum.

[MikeInnes]

Why would fill get a tuple of nothings as its gradient? That seems fishy and probably points to a different issue.

nothing (not a tuple) is a valid gradient, but the @adjoint macro should already be handling that.

[dsweber2]

here's a minimal example of what I was doing that caused the problem.

function example(x,N)
    ax = axes(x)
    extraAxe = ax[2+N:end]
    filledLoc = fill(1, N)
    return x[:, filledLoc..., extraAxe...]
end
y, back = pullback(example, randn(5,3,4,3), 2)
back(y)

which attempts to sum (nothing, nothing). Not sure where the code for it is, but it does make sense that the adjoint of splatting could result in multiple nothings grouped together.

[MikeInnes]

Ok, thanks for the clarification, I see what's happening. Unfortunately this is two things conspiring, the nothing gradient from indexing and the tuple-ing of those nothings from the splat (so easy to forget that iteration means you have to support tuples as a gradient). I think you already understood that but just so it's spelt out.

The best way to resolve this is probably to use accum_sum which handles nothings; I'll think about it bit I suspect we can't avoid passing the tuple of nothings here.

As an aside, this is quite a strange use of Fill; it would be more typical to use ntuple here since the compiled code will get more benefit from knowing the size of filledLoc rather than its particular value. This is not to say we shouldn't support it, of course.

[dsweber2]

it would be more typical to use ntuple

Good point, I should get more in the habit of using tuples. Thanks for looking into it.

[CarloLucibello]

closing as stale