NLSModels API

[tmigot]

Thanks @AntoninKns for this great repo!

I was checking the BundleAdjustmentModel, which follows the NLSModel API, and I have a couple of comments:

• I think the jacobian of the residual should be described by jac_structure_residual! and jac_coord_residual! and not jac_structure! L.159 and jac_coord! L.196.
• Adding the functions jprod_residual!, and jtprod_residual! would be very useful.
• I might have missed it, but it could be useful to add somewhere in the documentation that the BundleAdjustmentModel doesn't return any second-order information on the residual.
• Finally, I believe the docstring of BundleAdjustmentModel L.5 might be misleading since the model is a nonlinear least-square optimization problem.

[dpo]

@AntoninKns

[AntoninKns]

All issues have been fixed apart for the jprod_residual! and jtprod_residual!.

[AndrewLister-STFC]

I just tried to run your problems with the JSOSolvers and found I needed:

function NLPModels.jtprod_residual!(model::BundleAdjustmentModel{Float64, Vector{Float64}}, x, v, jv)
    mul!(jv, NLPModels.jac_residual(model, x)', v)
end
function NLPModels.jprod_residual!(model::BundleAdjustmentModel{Float64, Vector{Float64}}, x, v, jv)
    mul!(jv, NLPModels.jac_residual(model, x), v)
end

Not sure if you'd rather do something smarter but it's a quick workaround.

[tmigot]

Hi @AndrewLister-STFC ! Thank you for your interest and sorry for the delay of my response as we are understaffed at the moment.

You are right it is indeed a good workaround. I think long term we would have hope for a more efficient solution that would not evaluate (and therefore store) the Jacobian residual. My suggestion would be to combine this with ADNLPModel ( adapting this tuto https://jso.dev/ADNLPModels.jl/dev/mixed/ ) that will use AD for these operations.

Maybe this solution works with https://github.com/JuliaSmoothOptimizers/BundleAdjustmentModels.jl/pull/84 only.