kwargs for compute_extreme_point(lmo, direction)?

[gdalle]

For our use case in InferOpt.jl, we sometimes need to pass keyword arguments to the LMO. This is not allowed by FrankWolfe.jl, so I wrote a wrapper that predefines the kwargs with the LMO itself:

struct LMOWithKwargs{F,K} <: FrankWolfe.LinearMinimizationOracle
    minimizer::F
    minimizer_kwargs::K
end

LMOWithKwargs(minimizer) = LMOWithKwargs(minimizer, NamedTuple())

function FrankWolfe.compute_extreme_point(
    lmo::LMOWithKwargs, direction; kwargs...
)
    (; minimizer, minimizer_kwargs) = lmo
    v = minimizer(direction; minimizer_kwargs...)
    return v
end

Is there a more clever way to do this?

[matbesancon]

if you always pass the same kwargs to the LMO, why not set them as parameters of the LMO struct itself?

[gdalle]

That's exactly what I'm doing here (I edited the code above). But typically in InferOpt one might want to solve maximization problems on many different instances, passed as keyword arguments. So currently I re-create an LMO every time.
Basically I was wondering if something like this would make sense to you:

function frank_wolfe(..., lmo, ...; ..., lmo_kwargs, ...)
    for iteration in 1:n
        ...
        vertex = compute_extreme_point(lmo, direction, lmo_kwargs...)
        ...
    end
end