test_rrule leads to MethodError: Base.ArithmeticStyle(::Type{Union{}}) is ambiguous

[mcabbott]

I don't have a MWE, but the original case was here: https://github.com/JuliaDiff/ChainRules.jl/pull/570/files#diff-14d216cb53721d9360bbbcc399d93fff67f2594bd388162abbbe82c59ad1ba9fR236.

Maybe JuliaDiff/ChainRules.jl@944cced#diff-14d216cb53721d9360bbbcc399d93fff67f2594bd388162abbbe82c59ad1ba9fR236 is the right link.

[mzgubic]

Hmm, do you mean the error is

test_rrule: filter on Base.Fix2{typeof(>), Float64},Vector{Float64}: Error During Test at /Users/mzgubic/JuliaEnvs/ChainRules.jl/dev/ChainRulesTestUtils/src/testers.jl:195
  Got exception outside of a @test
  MethodError: no method matching zero(::Base.Fix2{typeof(>), Float64})
  Closest candidates are:
    zero(::Union{Type{P}, P}) where P<:Dates.Period at /Applications/Julia-1.7.app/Contents/Resources/julia/share/julia/stdlib/v1.7/Dates/src/periods.jl:53
    zero(::T) where T<:Dates.TimeType at /Applications/Julia-1.7.app/Contents/Resources/julia/share/julia/stdlib/v1.7/Dates/src/types.jl:450
    zero(::CartesianIndex{N}) where N at /Applications/Julia-1.7.app/Contents/Resources/julia/share/julia/base/multidimensional.jl:106
    ...

I think it comes from the fact that >(0.5) is essentially a closure over 0.5, and CRTU tries to wiggle it. I don't think this is what we want here, but in general I don't think we want to not wiggle closures, so adding the ⊢ NoTangent() is appropriate. IMO. Admittedly, the error message is bad, and I have no ideas on how to improve it.

[mcabbott]

Oh no, I thought I was linking to a commit, not just a line number.

No, Fix1 is another thing. Ideally it would perturb it, as you say, and deduce that it has no continuous effect, and thus accept Zero as the gradient.