Convex does not support broadcasting

[odow]

Convex.jl lacks support for a lot of broadcasting operations.

This causes a bunch of problems like this:

julia> using Convex

julia> x = Variable(2)
Variable
size: (2, 1)
sign: real
vexity: affine
id: 181…059

julia> A = rand(2, 2)
2×2 Matrix{Float64}:
 0.290045  0.201488
 0.890725  0.771457

julia> A * x .+ 1
ERROR: MethodError: no method matching iterate(::Convex.MultiplyAtom)
Closest candidates are:
  iterate(::Union{LinRange, StepRangeLen}) at range.jl:664
  iterate(::Union{LinRange, StepRangeLen}, ::Int64) at range.jl:664
  iterate(::T) where T<:Union{Base.KeySet{var"#s79", var"#s78"} where {var"#s79", var"#s78"<:Dict}, Base.ValueIterator{var"#s77"} where var"#s77"<:Dict} at dict.jl:693
  ...
Stacktrace:
 [1] copyto!(dest::Vector{Any}, src::Convex.MultiplyAtom)
   @ Base ./abstractarray.jl:843
 [2] _collect(cont::UnitRange{Int64}, itr::Convex.MultiplyAtom, #unused#::Base.HasEltype, isz::Base.HasLength)
   @ Base ./array.jl:608
 [3] collect(itr::Convex.MultiplyAtom)
   @ Base ./array.jl:602
 [4] broadcastable(x::Convex.MultiplyAtom)
   @ Base.Broadcast ./broadcast.jl:682
 [5] broadcasted(::Function, ::Convex.MultiplyAtom, ::Int64)
   @ Base.Broadcast ./broadcast.jl:1312
 [6] top-level scope
   @ REPL[4]:1

The purpose of this issue is to collate the currently open issues into a single thread, and to make it more discoverable for users coming to report bugs.

Replaces:

• Broadcasting issue with element-wise multiplication (Hadamard product) #410
• Element wise addition after multiplication #355
• Various operations don't support broadcast #218

[RoyiAvital]

In #717 a documentation warning was added.
It should be removed once this issue is solved.

[RoyiAvital]

I had another case with type conversion.
The following will fail:

function SolveSoftSVM1D( vX :: Vector{T}, vY :: Vector{N}, paramC :: T ) where {T <: AbstractFloat, N <: Integer}

    numPts = length(vX);

    if length(vY) != numPts
        throw(DimensionMismatch(lazy"`vY`` has length $(length(vY)), `vX` has length $(length(vX))"));
    end

    paramW = Variable();
    paramB = Variable();
    vξ     = Variable(numPts);

    sConvProb = minimize( T(0.5) * sumsquares(paramW) + paramC * sum(pos(one(T) - T.(vY) .* (paramW * vX + paramB))) );
    solve!(sConvProb, ECOS.Optimizer; silent = true);

    return paramW.value, paramB.value, vξ.value;
    
end

While changing T.(vY) into vY will work

[odow]

What's the full stack trace of the error?

[RoyiAvital]

Here is an MWE:

using Convex;
using ECOS;

function SolveSoftSVM1D( vX :: Vector{T}, vY :: Vector{N}, paramC :: T ) where {T <: AbstractFloat, N <: Integer}

    paramW = Variable();
    paramB = Variable();

    sConvProb = minimize( T(0.5) * sumsquares(paramW) + paramC * sum(pos(one(T) - T.(vY) .* (paramW * vX + paramB))) ); #<! Avoid casting with T.() in expression
    solve!(sConvProb, ECOS.Optimizer; silent = true);

    return paramW.value, paramB.value;
    
end

vX     = Float64.(collect(-3:3));
vY     = [-1, -1, -1, 1, 1, 1, 1];
paramC = 0.0;

paramW, paramB = SolveSoftSVM1D(vX, vY, paramC);

If you change T.(vY) into vY it will work.
The error:

ERROR: MethodError: no method matching iterate(::Convex.AdditionAtom)

Closest candidates are:
  iterate(::Base.AsyncGenerator, ::Base.AsyncGeneratorState)
   @ Base asyncmap.jl:362
  iterate(::Base.AsyncGenerator)
   @ Base asyncmap.jl:362
  iterate(::LibGit2.GitRevWalker)
   @ LibGit2 D:\Applications\JuliaWin\packages\julia\share\julia\stdlib\v1.10\LibGit2\src\walker.jl:29
  ...