I am trying to calculate the gradient of this function:
using Yota
f(x) = sum(getindex(x, [1 3; 4 2]))
s = rand(2,2)
val, g = Yota.grad(f, s)
If s = rand(2) my code works but the above code returns this error:
> Failed to find a derivative for %11 = hvcat(%6, %7, %8, %9, %10) ::Matrix{Int64} at position one
Why isn’t multidimensional indexing supported by default? Should I use @diffrule macro?
I edited my post.
What version of Julia are you using? I can’t run your function on Julia 1.5.0:
ulia> using Yota
julia> f(x) = sum(exp.(x), getindex(x, [1 3; 4 2]))
f (generic function with 1 method)
julia> s = rand(2,2)
2×2 Array{Float64,2}:
0.235071 0.271145
0.994342 0.97109
julia> f(x) = sum(exp.(x), getindex(x, [1 3; 4 2]))
f (generic function with 1 method)
julia> f(s)
ERROR: MethodError: objects of type Array{Float64,2} are not callable
Use square brackets [] for indexing an Array
Most likely this issue is not related to getindex(), but to sum(). I guess you try to use this method of sum():
sum(A::AbstractArray, w::StatsBase.UnitWeights; dims)
Which we don’t have a diff rule for. Is it the case?
No. If I use these functions I still see the same error:
f2(x) = sum(getindex(exp.(x), [1 3; 4 2]))
f3(x) = sum(exp.(getindex(x, [1 3; 4 2]))
Ok, with Julia 1.5.0 and Yota 0.4.4 I see another error related to the recent changes in getindex():
ERROR: BoundsError: attempt to access 1×2×2 Array{Float64,3} at index [1:1, 4, 1]
The changes were needed to support the case with repeating indices, e.g. getindex(x, [1, 2, 1]) (see related issue in Zygote if you’re curious about the whole story), but implementation is bit of a hack, so we get what we get.
I see 2 options for you here:
f2() and f3() on it and they work perfectly fine.