NLPModels

Knowledgeable users may have noticed that the Argos.AbstractNLPEvaluator API is closed to NLPModels. Hence, it is straightforward to wrap any AbstractNLPEvaluator in a NLPModels.AbstractNLPModel structure.

Initializing

In Argos, this is provided by the OPFModel structure, which takes as input any AbstractNLPEvaluator and converts it as an NLPModels.AbstractNLPModel`.

using NLPModels
using Argos

# Import OPF model in Argos
datafile = joinpath(INSTANCES_DIR, "case57.m")
flp = Argos.FullSpaceEvaluator(datafile)

# Convert it to a AbstractNLPModel:
model = Argos.OPFModel(flp)
@assert isa(model, NLPModels.AbstractNLPModel)

Playing with NLPModels' API

The user can use the standard NLPModels API to interact with the OPF model:

• Querying the number of variables:

julia> n = NLPModels.get_nvar(model)119

• Querying the initial variable x0:

julia> x0 = NLPModels.get_x0(model)119-element Vector{Float64}:
 -0.020594885173533087
 -0.10419615634406146
 -0.15097098029750955
 -0.07766715171374768
 -0.1668534764906579
 -0.18256143975860686
 -0.12775810124598497
 -0.14870205226991687
 -0.1322959573011702
 -0.19949113350295183
  ⋮
  1.005
  0.98
  1.015
  0.0
  0.4
  0.0
  4.5
  0.0
  3.1

• Evaluating the objective

julia> NLPModels.obj(model, x0)51272.87033037438

• Evaluating the constraints

julia> NLPModels.cons(model, x0)274-element Vector{Float64}:
  0.003022401147809717
  0.0030443460146013512
  0.0032197933906044085
  0.0033636219110988463
  0.002076154300374
  0.0001295163545154132
  0.0020171602977807623
 -0.003531140039454428
 -0.0030640689180172487
  0.0015899046184775223
  ⋮
  0.14159347915067355
  0.002641010671943805
  0.0031807469095941577
  0.0004765613725646054
  0.0021715752979237354
  0.0008227688607001616
  0.012748847855148328
  0.06416762549876907
  0.04544570155429488

• Evaluating the gradient

julia> NLPModels.grad(model, x0)119-element Vector{Float64}:
 -326563.85442535847
       0.0
       0.0
       0.0
       0.0
       0.0
       0.0
       0.0
       0.0
       0.0
       ⋮
       0.0
       0.0
       0.0
    4000.0
    4000.0
    4000.0
    3999.999998
    4000.0
    3999.999999

and so on...

Accelerating the callbacks on an NVIDIA GPU

We can exploit any available NVIDIA GPU to accelerate the evaluation of the derivatives. To do so, one first needs to install ArgosCUDA](../quickstart/cuda.md).

Then, we can instantiate a new evaluator on the GPU with:

using ArgosCUDA, CUDAKernels
flp = Argos.FullSpaceEvaluator(datafile; device=CUDADevice())

The OPFModel structure works exclusively on the host memory, so we have to bridge the evaluator flp to the host before creating a new instance of OPFModel:

brige = Argos.bridge(flp)
model = Argos.OPFModel(bridge)