error when trying to futher differentiate using Zygote

[btx0424]

Summary: it seems that the finite difference can't be further differentiated by Zygote Example: If I do

dif(x) = central_fdm(2, 1)(sin, x)

and then call

dif'(0)

I get the following error message:

setindex!(::StaticArrays.SVector{3, Float64}, value, ::Int) is not defined.

error(::String)@error.jl:33
setindex!(::StaticArrays.SVector{3, Float64}, ::Float64, ::Int64)@indexing.jl:3
#637@array.jl:318[inlined]
#2710#back@adjoint.jl:59[inlined]
Pullback@methods.jl:385[inlined]
(::typeof(∂(_estimate_magnitudes)))(::Tuple{Float64, Nothing})@interface2.jl:0
Pullback@methods.jl:365[inlined]
(::typeof(∂(estimate_step)))(::Tuple{Float64, Nothing})@interface2.jl:0
Pullback@methods.jl:193[inlined]
(::typeof(∂(λ)))(::Float64)@interface2.jl:0
Pullback@Other: 1[inlined]
(::typeof(∂(dif)))(::Float64)@interface2.jl:0
(::Zygote.var"#41#42"{typeof(∂(dif))})(::Float64)@interface.jl:41
gradient(::Function, ::Int64)@interface.jl:59
(::Zygote.var"#43#44"{typeof(Main.workspace50.dif)})(::Int64)@interface.jl:62
top-level scope@Local: 1[inlined]

The actual use case is that I'm trying to train a NN where I need to compute the gradient of its parameters w.r.t a loss function that involves finite differences given by central_fdm. Is this possible?

[simeonschaub]

Can you just use it the other way, i.e. FiniteDifferences over Zygote instead of Zygote over FiniteDifferences? You could actually define a custom adjoint for FiniteDifferenceMethod which does this automatically for cases like this.

[wesselb]

Step size adaptation appears to be not working well with Zygote. Unfortunately, I'm not familiar enough with the inner workings of Zygote to see what precisely breaks down. Perhaps @willtebbutt or @oxinabox could provide some insight. From the error message, if appears that a pullback attempts to modify a StaticArray in-place, which won't work.

Once you turn off step size adaptation, things seem to work:

julia> fdm = central_fdm(5, 1; adapt=0);

julia> cos_(x) = fdm(sin, x)
cos_ (generic function with 1 method)

julia> cos_'(1)
-0.8414709848079269

julia> -sin(1)
-0.8414709848078965

I would be careful with AD-ing through finite difference estimates, though. I second @simeonschaub's suggestion of taking finite differences of gradients computed by AD.

EDIT: If we wanted to make this work, we could do something like

using FiniteDifferences, Zygote

function (m::FiniteDifferences.AdaptedFiniteDifferenceMethod)(f::TF, x::Real) where TF<:Function
    x = float(x)  # Assume that converting to float is desired, if it isn't already.
    step = Zygote.dropgrad(first(FiniteDifferences.estimate_step(m, f, x)))
    return m(f, x, step)
end

Then

julia> fdm = central_fdm(5, 1; adapt=1);

julia> cos_(x) = fdm(sin, x)
cos_ (generic function with 1 method)

julia> cos_'(1)
-0.8414709848078701

julia> -sin(1)
-0.8414709848078965