Splatting and mutation?

[marcoct]

I ran into some unexpected nothings that depend on whether splatting was used or not.

I tried simplifying it into a minimal example---it seems like the problem might involve some combination of splatting and mutation.

import Zygote

mutable struct MyMutable
    value::Float64
end

function foo!(m::MyMutable, x)
    m.value = x
end

function baz(args)
    m = MyMutable(0.)
    # foo!(m, args[1]) # grad(1.) works as expected
    foo!(m, args...) # grad(1.) gives 'nothing'
    m.value
end

value, grad = Zygote.pullback(baz, (1.0,))
@assert grad(1.) == ((1.0,),)

PS: Zygote looks like great work, and we're looking forward to using it for AD within https://github.com/probcomp/Gen/

[DoktorMike]

I'm looking forward to this too as it will combine two of my favorite projects in julia. 😊 Will follow this!

[MikeInnes]

Your analysis is spot on -- Core._apply just assumed that f was pure, so didn't bother calling its adjoint with nothing as input. Thanks for the report and nice test case.

Glad to hear you're looking at this for Gen!