optimize lagrangian implementation in oop dispatch for autosparse backends

[baggepinnen]

σ = 0 is a common special case and it makes sense to optimize for it by not calling the cost function in this case

[ChrisRackauckas]

@gdalle @adrhill is there a way to force SCT to take the false path here?

[adrhill]

If lagrangian(θ, σ, λ, p) is a stable interface of OptimizationBase.jl, we could add a package extension to SCT to enable global sparsity patterns.

[baggepinnen]

lagrangian here is an inner function that calls other inner functions in the same scope.

[adrhill]

If multiple dispatch is an option, you could add a method that evaluates lagrangian(θ, σ, λ, p) = σ * f.f(θ, p) + dot(λ, cons_oop(θ)) on our global tracer types.

If MD is not an option, you could use the ifelse function, for which global sparsity detection will evaluate both branches.

[baggepinnen]

If MD is not an option, you could use the ifelse function, for which global sparsity detection will evaluate both branches.

Julia will also evaluate both branches, negating the optimization

[gdalle]

Is sigma related to the input here? Like, if x was a dual number, would sigma be one too? Or is it more of a fixed external parameter? Because depending on the answer, the if statement may cause issues with SparseConnectivityTracer

[baggepinnen]

not sure why I closed

Is sigma related to the input here? Like, if x was a dual number, would sigma be one too? Or is it more of a fixed external parameter?

It's up to the optimizer, so we cannot tell in general. I guess we could do something like

        function lagrangian(θ, σ, λ, p)
            if eltype(θ) <: SCT.TracerType || !iszero(θ)
                return σ * f.f(θ, p) + dot(λ, cons_oop(θ))
            else
                return dot(λ, cons_oop(θ))
            end
        end

[ChrisRackauckas]

That should be sufficient.

[Vaibhavdixit02]

This branch exists for most cases, but it looks like the oop case with autosparse was missed, so I have changed the title to reflect that.

[Vaibhavdixit02]

Do we need a compat bound on something based on the test failure?

[ChrisRackauckas]

The new title isn't correct, it's also optimizing the non-autosparse cases...

[ChrisRackauckas]

Do we need a compat bound on something based on the test failure?

No, we just need to keep pinging @vchuravy until the type piracy of Base goes away 😅

[Vaibhavdixit02]

The new title isn't correct, it's also optimizing the non-autosparse cases...

No
https://github.com/SciML/OptimizationBase.jl/pull/134/files#diff-7dfd1f414a8b2066bb9d673fb65b7fc09f9e695c5611bc7ff0ef534507ba0880

[ChrisRackauckas]

If you're pointing to this PR then yes it does the non sparse case...