Ambiguous Promotion Rule for <:Irrational,<:ReverseDiff.TrackedReal{V,D,O}}

[UserQuestions]

I was using ReverseDiff in a function that uses normcdf and observed the following error:

MethodError: promote_rule(::Type{Irrational{:invsqrt2}}, ::Type{ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}}}) is ambiguous. Candidates: promote_rule(::Type{R}, ::Type{ReverseDiff.TrackedReal{V,D,O}}) where {R<:Real, V, D, O} in ReverseDiff at HOMEDIR/.julia/v0.6/ReverseDiff/src/tracked.jl:257 promote_rule(::Type{#s268} where #s268<:Irrational, ::Type{T}) where T<:Number in Base at irrationals.jl:17 Possible fix, define promote_rule(::Type{R<:Irrational}, ::Type{ReverseDiff.TrackedReal{V,D,O}}) ForwardOptimize at macros.jl:126 [inlined]

• at scalars.jl:12 [inlined] normcdf(::ReverseDiff.TrackedReal{ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}},ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}},Void}) at norm.jl:15

I suspect that the corrective code to avoid ambiguity looks something like:

promote_rule{R<:Irrational, V, D, O}(::Type{R}, ::Type{ReverseDiff.TrackedReal{V,D,O}}) = TrackedReal{promote_type(R,V),D,O}

Am I correct? If so, is this line worth adding to the repository to avoid similar ambiguities with irrationals in the future?

[UserQuestions]

As a follow-up, it appears that in addition to the above, when creating tape to compute Hessians of functions with distributions (e.g. Distributions.normcdf), the Distributions package doesn't know how to handle the ReverseDiff.TrackedArray type. For example, when trying to apply ReverseDiff to create the [tape of] the Hessian for Distributions.normcdf, Julia is unable to match

normcdf(::ReverseDiff.TrackedArray{ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}},ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}},1,Array{ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}},1},Array{ReverseDiff.TrackedReal{Float64,Float64,ReverseDiff.TrackedArray{Float64,Float64,1,Array{Float64,1},Array{Float64,1}}},1}})

to either of the closest candidates:

normcdf(!Matched::Real, !Matched::Real, !Matched::Number) at HOME/.julia/v0.6/StatsFuns/src/distrs/norm.jl:16 normcdf(!Matched::Number) at HOME/.julia/v0.6/StatsFuns/src/distrs/norm.jl:15

[UserQuestions]

Bump

[jrevels]

For the OP, it might suffice just to add :Irrational to the list here.

For the second issue, that seems like a problem with whatever target code is calling that function, i.e. it looks external to ReverseDiff (I could be wrong, though). Why is a scalar function being called on an array?

[Moelf]

actually we should be able to just add AbstractIrrational to that list?

[ElOceanografo]

Just ran into this issue with ::Type{IrrationalConstants.Log2π}. Is there any reason not to add AbstractIrrational to REAL_TYPES? Happy to make a PR.