Slicing DenseAxisArrays

[IainNZ]

m = Model()
@defVar(m, x[0:5], Bin)
solve(m)
getValue(x[0])  # Works
getValue(x[0:2])
#BoundsError()
#while loading In[15], in expression starting on line 1
#
# in copy! at array.jl:52
# in getindex at array.jl:265
# in getindex at no file

[mlubin]

I think we don't transform the ranges, would be a good idea to do so...

[mlubin]

So this is working for JuMPArrays now. For JuMPDicts it's not immediately obvious what slice indexing for a dictionary should mean and how we can disambiguate it from using ranges as dictionary keys themselves.

[mlubin]

Actually that fix doesn't work because it breaks x[:], which is automatically translated by the compiler to x[1:length(x)]. Not sure how we can work around this...

[mlubin]

Which also reveals that we have no test coverage for the x[:] syntax.

[joehuchette]

Would adding a method with getindex(x::JuMPArray, c::Colon) work?

[mlubin]

I don't think : is currently passed through as a Colon.

[joehuchette]

Yeah, this is weird:

julia> x = rand(3,3);

julia> @which x[:]
getindex(A::Array{T,N},I::UnitRange{Int64}) at array.jl:262

[joehuchette]

Maybe worth opening an issue? Wanting to dispatch on Colon seems legitimate.

[IainNZ]

Yikes

[IainNZ]

https://groups.google.com/forum/#!topic/julia-dev/NOF6MA6tb9Y
I wonder if subarrays can do it somehow?

[IainNZ]

Ah I gues not, or rather, : does what they want anyway. I found the x[-1] behaviour a bit surprising!

julia> x = sub(A,2:3,2:3)
2x2 SubArray{Float64,2,Array{Float64,2},(UnitRange{Int64},UnitRange{Int64})}:
 0.0604057  0.953382
 0.0455991  0.919803

julia> x[:]
4-element Array{Float64,1}:
 0.0604057
 0.0455991
 0.953382
 0.919803

julia> x[1]
0.06040567704545796

julia> x[-1]
0.33475154171736476

[mlubin]

@timholy, are there plans to dispatch on Colon for indexing? I vaguely remember a discussion on this.

[timholy]

Yes, it seems likely to happen, but I wouldn't call it certain. There would be quite a few new methods needed for any object that supports indexing.

[joehuchette]

Looks like colon isn't desugared at parse time anymore (?):

julia> dump(:(a[1,:]))
Expr
  head: Symbol ref
  args: Array(Any,(3,))
    1: Symbol a
    2: Int64 1
    3: Symbol :
  typ: Any

[mlubin]

This has always been a compile-time transformation.

[mlubin]

The remaining issue, if any, is:

julia> m = Model();

julia> @variable(m, x[0:5]);

julia> x[0:2]
ERROR: KeyError: key 0:2 not found

[rapus95]

Are there any news on this? Why is slicing not supported in the first place?

[odow]

We're likely missing an appropriate getindex method. I don't think anyone has looked in detail about this recently.

[mlubin]

AxisArrays has this functionality in some form. You may be interested in https://github.com/rdeits/AxisArrayVariables.jl (which would need to be updated for the latest JuMP).

[rapus95]

How about sneakily passing it through?

JuMP.Containers.lookup_index(i, lookup::Dict) = isa(i, Colon) ? Colon() : isa(i, AbstractRange) ? i : lookup[i]

[mlubin]

What's the interpretation of indexing by 5:7 when the index set is 1:2:9? "passing it through" would be the equivalent of collect(1:2:9)[5:7] (BoundsError).

See the discussion at JuliaArrays/AxisArrays.jl#84 for more context.

[rapus95]

Better: if T::AbstractUnitRange then just pass through

[mlubin]

collect(5:7)[5:7] also yields a BoundsError.

[rapus95]

Is there any (dirty) workaround for making that

@variable(m, x[1:DN,1:DN,1:NTC,-36:T,1:NivelesC] >= 0, Int)
[constraints]
[optimize!]
JuMP.value.(x)[:, :, :, 3:40,:]

work easily?

[NLaws]

+1 on this issue