NLPModels computes gradient with elements out of order

[MaiconMares]

In some non linear problems the grad() function is computing gradient correctly but returning partial derivatives out of order. For example in the HS27 problem the correct result should be [16.02,-4.0,0.0], but it is returning [0.0,-4.0,16.02]. I don't understand why this behaviour is happening.

[dpo]

Is that with the JuMP model, or the ADNLPModel? Could you please give a snippet that reproduces the issue?

[MaiconMares]

Is that with the JuMP model, or the ADNLPModel? Could you please give a snippet that reproduces the issue?

@dpo I don't know what NLPModels uses under the hood to compute gradient, but to reproduce do it:

using NLPModels, CUTEst

nlp = CUTEstModel("HS27")
grad(nlp,nlp.meta.x0)

One screenshot with the result:

The initial values are [2.0,2.0,2.0].

[MaiconMares]

@dpo this behaviour is strange because in some NLP problems the gradient is computed in the correct order, I mean:

[dpo]

Ok, you're in the wrong repository. You're talking about CUTEst.jl models.

[abelsiqueira]

Hi @MaiconMares, what order were you expecting? The paper's or wrt other packages? Maybe it is related to the setup options that we use to read the cutest file.

[MaiconMares]

Hi @MaiconMares, what order were you expecting? The paper's or wrt other packages? Maybe it is related to the setup options that we use to read the cutest file.

Hi @abelsiqueira the correct order should be [16.02,-4.0,0.0]. How can I change the way CUTEst Julia package reads .SIF files?

[abelsiqueira]

CUTEst.jl/src/CUTEst.jl

Line 264 in 9467c8c

function CUTEstModel(

See above and check the docs of the function.

[tmigot]

Hey @MaiconMares , what is reference you use to state what is the ''correct'' formulation?
Or is an error on the derivative computation in cutest?

[MaiconMares]

Hi @MaiconMares, what order were you expecting? The paper's or wrt other packages? Maybe it is related to the setup options that we use to read the cutest file.

Thank you @abelsiqueira, it was the problem. lfirst and lvfirst args passed to CUTEstModel function was leading CUTEst to compute gradients and constraints doing partial derivatives over non-linear variables first! The solution:

CUTEstModel("HS8",lfirst=false,lvfirst=false)