Can't differentiate foreigncall expression when trying to compute gradient

[LucasPa]

Hi,

I am beginner in Julia and I would like to use Flux. I tried to do the following:

using Flux
NN = Chain(Dense(1, 5, σ), Dense(5, 1, σ))

model(x) = NN([x])[1];
df(x) = gradient(model, x)[1]

function lossEq(x)
(df(x) - 2.0 * x - 3)^2
end

gs = gradient(() -> lossEq(0.2), params(NN))

But the following error is raised:

ERROR: Can't differentiate foreigncall expression

Is it actually possible to compute such a gradient that involves a gradient ?

Best,

Lucas

[DhairyaLGandhi]

You should be able to see this. You could also check out https://github.com/DhairyaLGandhi/Zygote.jl/tree/dg/iddict to see if that works out better

[DhairyaLGandhi]

Also, when taking nested gradients, it might be better to call y, back = Zygote.pullback(...) and then back(...) explicitly.