Array of Variable is Unexpected behavior

[utotch]

Hello.
I found that I can't make "Array of Variable" by using Convex.jl.
I think it is unexpected behavior but is it a supposed specification?

I think a collection of Variable is useful in some situations and this behavior limits my coding style...

-- simple example

using Convex
a = Variable(2)
[a]

MethodError: no method matching +(::Convex.NoSign)
Closest candidates are:
  +(::Convex.NoSign, ::Convex.NoSign) at /opt/julia/v0.6/Convex/src/dcp.jl:93
  +(::Convex.NoSign, ::Convex.Positive) at /opt/julia/v0.6/Convex/src/dcp.jl:94
  +(::Convex.NoSign, ::Convex.Negative) at /opt/julia/v0.6/Convex/src/dcp.jl:96
  ...

Stacktrace:
 [1] sum(::Tuple{Convex.NoSign}) at ./tuple.jl:323
 [2] sign(::Convex.HcatAtom) at /opt/julia/v0.6/Convex/src/atoms/affine/stack.jl:26
 [3] show(::IOContext{Base.AbstractIOBuffer{Array{UInt8,1}}}, ::Convex.TransposeAtom) at /opt/julia/v0.6/Convex/src/utilities/show.jl:56
 [4] limitstringmime(::MIME{Symbol("text/plain")}, ::Convex.TransposeAtom) at /opt/julia/v0.6/IJulia/src/inline.jl:24
 [5] display_dict(::Convex.TransposeAtom) at /opt/julia/v0.6/IJulia/src/execute_request.jl:28
 [6] (::Compat.#inner#17{Array{Any,1},IJulia.#display_dict,Tuple{Convex.TransposeAtom}})() at /opt/julia/v0.6/Compat/src/Compat.jl:488
 [7] execute_request(::ZMQ.Socket, ::IJulia.Msg) at /opt/julia/v0.6/IJulia/src/execute_request.jl:186
 [8] (::Compat.#inner#17{Array{Any,1},IJulia.#execute_request,Tuple{ZMQ.Socket,IJulia.Msg}})() at /opt/julia/v0.6/Compat/src/Compat.jl:488
 [9] eventloop(::ZMQ.Socket) at /opt/julia/v0.6/IJulia/src/eventloop.jl:8
 [10] (::IJulia.##14#17)() at ./task.jl:335

-- and can't make Array of Variable

using Convex
A = [Variable(2), Variable(3,4)]

Cannot horizontally stack expressions of varying number of rows

Stacktrace:
 [1] Convex.HcatAtom(::Convex.Variable, ::Convex.Variable) at /opt/julia/v0.6/Convex/src/atoms/affine/stack.jl:16
 [2] hcat(::Convex.Variable, ::Convex.Variable) at /opt/julia/v0.6/Convex/src/atoms/affine/stack.jl:113
 [3] include_string(::String, ::String) at ./loading.jl:522

[utotch]

I found that "append!" don't work but "push!" works well.

[Non-Contradiction]

Maybe you can use other structures to hold the collection of variables instead of array, like tuple or dict?

As far as I can see, the array or variables (when it works) may not behave like what you want.
For example, [Variable(2), Variable(3)][1] is a Variable(1), not Variable(2) you may expect.
I think it is because variables are treated like vectors/matrices, and [] is overloaded to concatenate the vectors/matrices. So A = [Variable(2), Variable(3,4)] error message is
"Cannot horizontally stack expressions of varying number of rows"

[utotch]

I implemented my problem without using append!.
I prepare empty array [] and push! Variables.

Thus I was able to use Array for a collection of Variables, but I think overriding [] is confusing.

[ericphanson]

Note that

a = Variable(2)
[a] 

works fine in the current version of Convex. Regarding

A = [Variable(2), Variable(3,4)]

the problem seems to be these lines: https://github.com/JuliaOpt/Convex.jl/blob/fe366821b77d8243f2ca21a7deb35f4d50c579ce/src/atoms/affine/stack.jl#L135-L139

What's interesting is that Convex.jl's tests pass when I just comment them out, and moreover

A = [Variable(2), Variable(3,4)]

works fine.

Maybe we can just remove those problematic lines? I see they were introduced years ago in a much older version of Julia, and maybe they are no longer necessary. I will submit a pull request to remove the Base.vect override for discussion.