Categorical cannot take abstract array p

[cossio]

For example:

P = abs.(randn(5,4,2));
p = view(P,:,1,1);
p ./=sum(p);

Then:

julia> Categorical(p)
ERROR: MethodError: Cannot `convert` an object of type Array{Float64,1} to an object of type SubArray{Float64,1,Array{Float64,3},Tuple{Base.Slice{Base.OneTo{Int64}},Int64,Int64},true}
Closest candidates are:
  convert(::Type{T}, ::T) where T<:AbstractArray at abstractarray.jl:14
  convert(::Type{T}, ::LinearAlgebra.Factorization) where T<:AbstractArray at /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.3/LinearAlgebra/src/factorization.jl:53
  convert(::Type{T}, ::T) where T at essentials.jl:168
  ...
Stacktrace:
 [1] (::Distributions.var"#_#23#24")(::Bool, ::Type{DiscreteNonParametric{Int64,Float64,Base.OneTo{Int64},SubArray{Float64,1,Array{Float64,3},Tuple{Base.Slice{Base.OneTo{Int64}},Int64,Int64},true}}}, ::Base.OneTo{Int64}, ::SubArray{Float64,1,Array{Float64,3},Tuple{Base.Slice{Base.OneTo{Int64}},Int64,Int64},true}) at /home/cossio/.julia/packages/Distributions/kPXE9/src/univariate/discrete/discretenonparametric.jl:31
 [2] Type at ./none:0 [inlined]
 [3] #_#26 at /home/cossio/.julia/packages/Distributions/kPXE9/src/univariate/discrete/categorical.jl:31 [inlined]
 [4] Type at ./none:0 [inlined]
 [5] #Categorical#27 at /home/cossio/.julia/packages/Distributions/kPXE9/src/univariate/discrete/categorical.jl:34 [inlined]
 [6] DiscreteNonParametric{Int64,P,Base.OneTo{Int64},Ps} where Ps where P(::SubArray{Float64,1,Array{Float64,3},Tuple{Base.Slice{Base.OneTo{Int64}},Int64,Int64},true}) at /home/cossio/.julia/packages/Distributions/kPXE9/src/univariate/discrete/categorical.jl:34
 [7] top-level scope at REPL[16]:1

Cannot Categorical(p) admit generic AbstractVector as argument p?

[findmyway]

You can bypass the check with Categorical(p;check_args=false)

[baedan]

+1 to this. the error is hard to make sense of also

[vandenman]

Just got bitten by this... The issue is that in the constructor for DiscreteNonParametric there is

Distributions.jl/src/univariate/discrete/discretenonparametric.jl

Lines 33 to 34 in a1010e4

	sort_order = sortperm(xs)
	new{T,P,Ts,Ps}(xs[sort_order], ps[sort_order])

The return type of ps[sort_order] is not a view/ subarray but rather a vector. However, the type parameter Ps is inferred from the originally passed ps, which leads to the convert error.

I don't think it's a good idea to make Ps depend on the result of ps[sort_order], because then the constructor becomes type unstable (the branch that does not check arguments would return a different type for views).

IMO a better approach here is to ensure that the constructor for Categorical never does sort_order = sortperm(xs) because

Distributions.jl/src/univariate/discrete/categorical.jl

Line 27 in a1010e4

const Categorical{P<:Real,Ps<:AbstractVector{P}} = DiscreteNonParametric{Int,P,Base.OneTo{Int},Ps}

implies that xs is always sorted. So perhaps it makes sense to add a specialized constructor for DiscreteNonParametric for T<:Integer and Ts<:Base.OneTo{<:Integer} that skips the sorting step. Perhaps something along the lines of

 function DiscreteNonParametric{T,P,Ts,Ps}(xs::Ts, ps::Ps; check_args::Bool=true) where {
            T<:Integer,P<:Real,Ts<:Base.OneTo{T},Ps<:AbstractVector{P}}
        check_args || return DiscreteNonParametric{T,P,Ts,Ps}(xs, ps, check_args = false)
        @check_args(
            DiscreteNonParametric,
            (length(xs) == length(ps), "length of support and probability vector must be equal"),
            (ps, isprobvec(ps), "vector is not a probability vector")
        )
        return DiscreteNonParametric{T,P,Ts,Ps}(xs, ps, check_args = false)
    end

I'd be happy to open a PR!