SparseArrays.jl

SparseArrays.jl copied to clipboard

Julia is not detecting reshaped sparse arrays as sparse. This has a wierd consequence when trying to create a sparse vector from a sparse matrix.

julia> D = sparse(Diagonal([1,1,1,1]))
4×4 SparseMatrixCSC{Int64,Int64} with 4 stored entries:
  [1, 1]  =  1
  [2, 2]  =  1
  [3, 3]  =  1
  [4, 4]  =  1

julia> v = sparsevec(D)
16-element reshape(::SparseMatrixCSC{Int64,Int64}, 16) with eltype Int64:
 1
 0
 0
 0
 0
 1
 0
 0
 0
 0
 1
 0
 0
 0
 0
 1

julia> issparse(v)
false

julia> findnz(v)
ERROR: MethodError: no method matching findnz(::Base.ReshapedArray{Int64,1,SparseMatrixCSC{Int64,Int64},Tuple{Base.MultiplicativeInverses.SignedMultiplicativeInverse{Int64}}})
Closest candidates are:
  findnz(::SparseArrays.AbstractSparseMatrixCSC{Tv,Ti}) where {Tv, Ti} at C:\buildbot\worker\package_win64\build\usr\share\julia\stdlib\v1.5\SparseArrays\src\sparsematrix.jl:1453
  findnz(::SparseVector{Tv,Ti}) where {Tv, Ti} at C:\buildbot\worker\package_win64\build\usr\share\julia\stdlib\v1.5\SparseArrays\src\sparsevector.jl:735
Stacktrace:
 [1] top-level scope at REPL[116]:1

I think the best way to fix this would be to add a ReshapedSparseArray <: AbstractSparseArray wrapper type. This type could then be used not only for sparsevec but also for reshapeing sparse arrays in general.

Maybe the sparsevec definition could be changed to something like

sparsevec(a::AbstractSparseArray) = sparsevec(vec(a))

?

That would be a simple solution, but seems inefficient. It would also create a completely new array that shares no memory with the original one, in contrast to the current behavior where modifying the new array will also modify the old one.