test fails on Julia 1.8

[Moelf]

using Integrals, Zygote

julia> f(x, p) = sum(sin.(p[1] * x))
f (generic function with 1 method)

julia> lb = 1.0
1.0

julia> ub = 3.0
3.0

julia> p = 2.0
2.0

julia> function testf(lb, ub, p)
           prob = IntegralProblem(f, lb, ub, p)
           sin(solve(prob, QuadGKJL(), reltol = 1e-3, abstol = 1e-3)[1])
       end
testf (generic function with 1 method)

julia> dlb1, dub1, dp1 = Zygote.gradient(testf, lb, ub, p)
ERROR: BoundsError: attempt to access 0-element Vector{Any} at index []
Stacktrace:
  [1] throw_boundserror(A::Vector{Any}, I::Tuple{})
    @ Base ./abstractarray.jl:703

[ChrisRackauckas]

On v1.8 with the latest Integrals this outputs:

julia > dlb1, dub1, dp1 = Zygote.gradient(testf, lb, ub, p)
(-0.9092974268256817, -0.27941549819892586, -0.4091435516769757)

[nelslind]

I'm having the same issue in Julia 1.8.5. What version of Julia and versions of Zygote and Integrals was this working under?

[ChrisRackauckas]

The example is incorrect. Corrected:

using Integrals, Zygote
f(x, p) = sum(sin.(p * x))
lb = 1.0
ub = 3.0
p = 2.0
function testf(lb, ub, p)
    prob = IntegralProblem(f, lb, ub, p)
    sin(solve(prob, QuadGKJL(), reltol = 1e-3, abstol = 1e-3)[1])
end
dlb1, dub1, dp1 = Zygote.gradient(testf, lb, ub, p)

The difference is f(x, p) = sum(sin.(p * x)) instead of f(x, p) = sum(sin.(p[1] * x)). The OP's original example fails because p[1] fails because p is a scalar, not a vector. I don't know why Zygote throws that weird error, but that's an issue for Zygote.

If you have an example that' supposed to work, please open a new issue.

[Moelf]

? that's the problem though

julia> p = [2.0]
1-element Vector{Float64}:
 2.0

julia> function testf(lb, ub, p)
           prob = IntegralProblem(f, lb, ub, p)
           sin(solve(prob, QuadGKJL(), reltol = 1e-3, abstol = 1e-3)[1])
       end
testf (generic function with 1 method)

julia> dlb1, dub1, dp1 = Zygote.gradient(testf, lb, ub, p)
ERROR: BoundsError: attempt to access 0-element Vector{Any} at index []
Stacktrace:
  [1] throw_boundserror(A::Vector{Any}, I::Tuple{})
    @ Base ./abstractarray.jl:744
  [2] checkbounds
    @ ./abstractarray.jl:709 [inlined]
  [3] _getindex
    @ ./abstractarray.jl:1326 [inlined]
  [4] getindex
    @ ./abstractarray.jl:1294 [inlined]
  [5] macro expansion
    @ ~/.julia/packages/Zygote/SuKWp/src/lib/lib.jl:314 [inlined]
  [6] (::Zygote.Jnew{Base.MethodList, Vector{Any}, false})(Δ::Nothing)
    @ Zygote ~/.julia/packages/Zygote/SuKWp/src/lib/lib.jl:308

also Zygote can handle this:

julia> p = [2.0]
1-element Vector{Float64}:
 2.0

julia> Zygote.gradient(f, 30, p)
(-1.9048259608303126, [-28.57238941245469])

julia> p = 2.0
2.0

julia> Zygote.gradient(f, 30, p)
(-1.9048259608303126, -28.57238941245469)

[nelslind]

Hi All, Thanks for the help with this.

Here is a simple example:

`
foo(x, p) = p + x

lb = 1.0
ub = 3.0
p = 2.0

function testf(lb, ub, p)
prob = IntegralProblem(foo, lb, ub, p)
solve(prob, QuadGKJL(), reltol = 1e-3, abstol = 1e-3)[1]
end

dlb1, dub1, dp1 = Zygote.gradient(testf, lb, ub, p)
`

I get "ERROR: BoundsError: attempt to access 0-element Vector{Any} at index []"

[nelslind]

Also, the example at

https://docs.juliahub.com/Integrals/QIwht/3.1.1/tutorials/differentiating_integrals/

creates the same error for Zygote, but works fine for ForwardDiff and and FiniteDiff (I had to replace CubaCuhre with HCubatureJL). This is true using either Integrals 3.1.1 (which is the version listed at that page) and 3.7.

[ChrisRackauckas]

@ArnoStrouwen does that already have an issue?

[ArnoStrouwen]

I think it is this issue #99?