SparseMatrixCSC does not support indexing by arrays of CartesianIndex

[yurivish]

With Julia 1.1:

julia> using SparseArrays

julia> let r = [1, 2, 3]
           r[findall(r .> 2)]
       end
1-element Array{Int64,1}:
 3

julia> let r = sparse([1, 2, 3])
           r[findall(r .> 2)]
       end
1-element SparseVector{Int64,Int64} with 1 stored entry:
  [1]  =  3

julia> let r = [1 2; 3 4]
           r[findall(r .> 2)]
       end
2-element Array{Int64,1}:
 3
 4

julia> let r = sparse([1 2; 3 4])
           r[findall(r .> 2)]
       end
ERROR: MethodError: no method matching isless(::CartesianIndex{2}, ::Int64)
Closest candidates are:
  isless(::Missing, ::Any) at missing.jl:70
  isless(::AbstractFloat, ::Real) at operators.jl:150
  isless(::Real, ::Real) at operators.jl:338
  ...
Stacktrace:
 [1] <(::CartesianIndex{2}, ::Int64) at ./operators.jl:260
 [2] getindex(::SparseMatrixCSC{Int64,Int64}, ::Array{CartesianIndex{2},1}) at /Users/osx/buildbot/slave/package_osx64/build/usr/share/julia/stdlib/v1.1/SparseArrays/src/sparsevector.jl:669
 [3] top-level scope at REPL[52]:2

[nicrummel]

I also am experiencing this issue

[StefanKarpinski]

Is this a bug or a missing feature?

[nicrummel]

Is this a bug or a missing feature?

I think this is a missing feature. Arrays can be accessed at multiple Cartesian Indices through an array, CartesianIndex[], to obtain the same results we would have to use a list comprehension with sparse arrays.

Example Code:

idx1 = CartesianIndex(1,2)
idx2 = CartesianIndex(3,4)
a = zeros(5,5)
a[idx1] = rand()
a[idx2] = rand()
@show a[[idx1,idx2]]
import SparseArrays
b = SparseArrays.SparseMatrixCSC(a)
@show b[[idx1,idx2]]
@show [b[i] for i in [idx1,idx2]]

Output:

[mbauman]

I still lean towards bug because this should work for all array subtypes but no matter.

[yurivish]

Yeah, I encountered it as a bug.

In my mind, there is a contract to the array interface by which any dimension can be indexed by an array of elements, each of which is a valid index.

[nicrummel]

I believe this would fix the bug/enhancement ...

[mbauman]

These indexing definitions are fairly old — there could be some nice gains to be had by using some of the new tools we've developed over the past four+ years. In this case, it'd just be simpler (and faster in some cases) to just do a checkbounds(A, I) at the top of the function.