Nicer workflow for non-Float64

[jeremypparker]

Thanks for all the work on v0.8

This now works:

using SumOfSquares
using DynamicPolynomials
using Hypatia
using DoubleFloats


model = GenericModel{Double64}(Hypatia.Optimizer{Double64})

@polyvar x

_1 = one(GenericAffExpr{Double64, GenericVariableRef{Double64}})

p = _1 + x^2

@constraint(model, p in SOSCone())

optimize!(model)

However, it's very fragile and fiddly: even replacing using _1 = one(Double64) causes an error. I guess there could be GenericSOSModel very easily, I'm more concerned about the issue of p having hidden Float64 inside.

Setting a domain in the constraint does not work either

[jeremypparker]

I got domain working but it's very ugly:

using SumOfSquares
using DynamicPolynomials
using Hypatia
using DoubleFloats

using MathOptInterface

const MOI = MathOptInterface

function _to_Double64(f::MOI.ScalarAffineFunction{Float64})
    return MOI.ScalarAffineFunction{Double64}(
        MOI.ScalarAffineTerm{Double64}[
            MOI.ScalarAffineTerm{Double64}(Double64(t.coefficient), t.variable)
            for t in f.terms
        ],
        Double64(f.constant),
    )
end

import Base: /
/(f::MOI.ScalarAffineFunction{Float64}, a::Double64) = _to_Double64(f) / a

model = GenericModel{Double64}(Hypatia.Optimizer{Double64})

@polyvar s c

_1 = one(GenericAffExpr{Double64, GenericVariableRef{Double64}})

dom = @set c^2 + s^2 - 1 == 0

p = _1 - s^2

cref = @constraint(model, p in SOSCone(), domain = dom)


optimize!(model)

[blegat]

Good point, thanks for raising this. In JuMP, we just convert the coefficients of what's in @constraint into the coefficient type of the GenericModel so we should do this here too. I'll take a look