AutoFastDifferentiation not working with NamedTuple context

[pitx-perf]

Issue

AutoFastDifferentiation is not able to deal with a NamedTuple context which could happen when the parameters of an optimization is a NamedTuple.

The error message:

ERROR: MethodError: no method matching variablize(::@NamedTuple{v1::Float64, v2::Float64}, ::Symbol) 
Closest candidates are:
  variablize(::AbstractArray, ::Symbol)
  variablize(::Number, ::Symbol)

MWE

Example taken from Optimization.jl, and modified to have a NamedTuple as p and use AutoFastDifferentiation

using Optimization
import FastDifferentiation

rosenbrock(u, p) = (p.v1 - u[1])^2 + p.v2 * (u[2] - u[1]^2)^2
u0 = zeros(2)
p = (v1=1.0, v2=100.0)

optf = OptimizationFunction(rosenbrock, AutoFastDifferentiation())
prob = OptimizationProblem(optf, u0, p)

sol = solve(prob, Optimization.LBFGS())

AutoForwardDiff works fine.

[gdalle]

Hi @pitx-perf,
Thanks for reporting this. Indeed, AutoFastDifferentiation is one of those backends which require any Constant or Cache to be either a Number or an AbstractArray (as hinted in the Constant docstring for example). I just haven't had time to implement more flexibility, because I didn't think anyone was actually using the symbolic backends in sophisticated way. Do you want to try and contribute a fix? I'd be happy to guide you, and it's good for me to get more people familiar with the DI code.

[pitx-perf]

Hi @gdalle,
Yes I'd be happy to contribute to a fix. I definitely need some time to get familiar with both DI and FastDifferentiation beforehand.
I faced this issue just trying several backends. No idea if FastDifferentiation is actually suited for my problem.

[gdalle]

No pressure of course, just let me know if you end up not wanting to write the fix and I'll take care of it.
Out of curiosity, what is your problem?