Softmax segfaults with offset arrays

[juliastats]

Softmax segfaults with offset arrays

[mcabbott]

BTW, the quick-and-dirty "fix" is to copy LinearAlgebra and insert this all over the place:

julia> LinearAlgebra.require_one_based_indexing(zeros(3))
true

julia> LinearAlgebra.require_one_based_indexing(zeros(0:2))
ERROR: ArgumentError: offset arrays are not supported but got an array with index other than 1

(Only exists on Julia 1.2 & up, obviously trivial to roll your own.) Of course better to fix things to allow offsets, but not a bad first step to forbid them.

[yurivish]

Agreed, using Base.require_one_based_indexing would be an improvement over the current state of affairs. It's documented in the devdocs, though as you say only available on Julia 1.2 and up.

What is zeros(0:2)? On Julia 1.5.3 and 1.6 I get an error:

julia> zeros(0:2)
ERROR: MethodError: no method matching zeros(::Type{Float64}, ::Tuple{UnitRange{Int64}})
Closest candidates are:
  zeros(::Type{T}, ::Union{Integer, AbstractUnitRange}...) where T at array.jl:521
  zeros(::Type{T}, ::Tuple{}) where T at array.jl:529
  zeros(::Type{T}, ::Tuple{Vararg{Integer,N}}) where {T, N} at array.jl:524
  ...
Stacktrace:
 [1] zeros(::Tuple{UnitRange{Int64}}) at ./array.jl:522
 [2] zeros(::UnitRange{Int64}) at ./array.jl:520
 [3] top-level scope at REPL[13]:1
``

[andreasnoack]

I think it would be almost as simple to just fix all the loops to loop over the values instead of the indices.