Pass sparse structure of constraints to NLPModels

[jbcaillau]

See control-toolbox/CTBenchmarks.jl#17 (comment)

[jbcaillau]

@amontoison can you please point towards a worked out example defining an NLPModel? According to the API, I guess that the stuff listed here are required (Jacobian / Hessian and associated products...)

[jbcaillau]

See JuliaSmoothOptimizers/ADNLPModels.jl#284

[amontoison]

Do you want a ManualNLPModel or still use an ADNLPModel?
-> https://jso.dev/ADNLPModels.jl/previews/PR209/sparsepattern/

[jbcaillau]

Thanks @amontoison for the example. I suppose that building manually the whole thing (= sparse derivatives, using AD only where it’s needed) gives the best result, but what is tour advice?

[amontoison]

@jbcaillau If you directly provide the sparsity pattern of Jacobians / Hessians in an ADNLPModel / ADNLSModel, you will have a speed-up for sure.

If you are building manually the whole thing, the coloring / AD / decompression will be also replaced by your own function jac_coord and hess_coord and it highly depends on your implementation of these fonctions.
If you exploit the structure and redondant patterns, you will have an advantage but you also need to find some parallelism.
(It's a good application for Tapenade with code to code AD.)

With AD we can evaluate multiple directional derivates in the same time to compute the compressed Jacobian / Hessian.

[jbcaillau]

@jbcaillau If you directly provide the sparsity pattern of Jacobians / Hessians in an ADNLPModel / ADNLSModel, you will have a speed-up for sure.

If your building manually the whole thing, the coloring / AD / decompression will be also replaced by your own function jac_coord and hess_coord and it highly depends on your implementation of these fonctions […]

thanks @amontoison let’s try the first option as a start. looking forward JuliaSmoothOptimizers/ADNLPModels.jl#286

[jbcaillau]

@amontoison actually, there is a worked out example of building an NLPModel by hand in MadNLP docs:
https://madnlp.github.io/MadNLP.jl/dev/quickstart/#Specifying-the-derivatives-by-hand:-NLPModels.jl

[amontoison]

I already gave you a tutorial on a ManualNLPModel, is it not what you wanted?
https://jso.dev/tutorials/create-a-manual-model/

In the same spirit than the example of MadNLP, we have a few problems defined by hand to test our various AbstractNLPModels in the JSO ecosystem:
https://github.com/JuliaSmoothOptimizers/NLPModelsTest.jl/tree/main/src/nlp/problems