LoadError: ArgumentError: Converting an instance of ReverseDiff.TrackedReal{Float64, Float64, Nothing} to Float64 is not defined. Please use `ReverseDiff.value` instead

General Usage
,
1

Hi,

I’d like to take a gradient of a simple function using ReverseDiff, however I stumbled upon an error which I haven’t been able figure out, the code which throws the error is

using ReverseDiff

function f_V(R, a, b, c, d, g)
    V = 0.
    if R ≤ b
        V = Inf
    elseif b ≤ R ≤ c
        R2 = R^2
        V = -a*(c^2 - R2)^g * ((R2 - d^2)/(R2 - b^2))
    end
    return V
end

function f_U(R, indexer, arg_v...)
    n_data = size(R)[1]; n_dim = size(indexer)[2]
    Vref = f_V.(R, arg_v...)
    U = Matrix{Float64}(undef, n_data, n_dim)
    @simd for i=1:n_dim
        Vsub = @view Vref[:, indexer[:,i]]
        U[:, i] = sum(Vsub, dims=2)
    end
    U = U./maximum(abs.(U))
    return U
end

R = rand(10, 3)
a = 1.; b = 1e-6; c = 2.; d = .9; g = 6. 
indexer = [1 1 2; 2 3 3]

ReverseDiff.gradient(var -> f_U(R, indexer, var[1], b, c, d, g), [a])

The complete error trace is

ERROR: LoadError: ArgumentError: Converting an instance of ReverseDiff.TrackedReal{Float64, Float64, Nothing} to Float64 is not defined. Please use `ReverseDiff.value` instead.
Stacktrace:
  [1] convert(#unused#::Type{Float64}, t::ReverseDiff.TrackedReal{Float64, Float64, Nothing})
    @ ReverseDiff C:\Users\beryl\.julia\packages\ReverseDiff\Y5qec\src\tracked.jl:261
  [2] setindex!
    @ .\array.jl:905 [inlined]
  [3] macro expansion
    @ .\multidimensional.jl:910 [inlined]
  [4] macro expansion
    @ .\cartesian.jl:64 [inlined]
  [5] _unsafe_setindex!(::IndexLinear, ::Matrix{Float64}, ::Matrix{ReverseDiff.TrackedReal{Float64, Float64, Nothing}}, ::Base.Slice{Base.OneTo{Int64}}, ::Int64)
    @ Base .\multidimensional.jl:905
  [6] _setindex!
    @ .\multidimensional.jl:894 [inlined]
  [7] setindex!
    @ .\abstractarray.jl:1315 [inlined]
  [8] macro expansion
    @ C:\Users\beryl\Documents\Coding\Python\pes\test.jl:20 [inlined]
  [9] macro expansion
    @ .\simdloop.jl:77 [inlined]
 [10] f_U(::Matrix{Float64}, ::Matrix{Int64}, ::ReverseDiff.TrackedReal{Float64, Float64, ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}}, ::Vararg{Any})
    @ Main C:\Users\beryl\Documents\Coding\Python\pes\test.jl:18
 [11] (::var"#5#6")(var::ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}})
    @ Main C:\Users\beryl\Documents\Coding\Python\pes\test.jl:30
 [12] ReverseDiff.GradientTape(f::var"#5#6", input::Vector{Float64}, cfg::ReverseDiff.GradientConfig{ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}})
    @ ReverseDiff C:\Users\beryl\.julia\packages\ReverseDiff\Y5qec\src\api\tape.jl:199
 [13] gradient(f::Function, input::Vector{Float64}, cfg::ReverseDiff.GradientConfig{ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}}) (repeats 2 times)
    @ ReverseDiff C:\Users\beryl\.julia\packages\ReverseDiff\Y5qec\src\api\gradients.jl:22
 [14] top-level scope
    @ C:\Users\beryl\Documents\Coding\Python\pes\test.jl:30
 [15] include(fname::String)
    @ Base.MainInclude .\client.jl:451
 [16] top-level scope
    @ REPL[1]:1
in expression starting at C:\Users\beryl\Documents\Coding\Python\pes\test.jl:30

The only thing that worked was when I took the gradient from f_V directly since technically the only function that is differentiated is f_V, as follows

ReverseDiff.gradient(var -> f_V(R[1,1], var[1], b, c, d, g), [a]) #<<--- works correctly

However I’d like to do the gradient routine from f_U (actually, from a function which calls f_U, I tried to narrow down the error it seems like it’s coming from f_U), since it’ll be used as a derivative to be fed into Optim with [a,b,c,d] as (the subset of) the tuning parameters. Thus I’d like to solve this error. Thanks.