Support the JuMP macro operator

src/MOI_wrapper.jl

@@ -61,6 +61,7 @@ function MOI.supports(
     MOI.ObjectiveSense,
     MOI.ObjectiveFunction{<:Union{LinQuad, MOI.ScalarNonlinearFunction}},
     MOI.NLPBlock,
+    MOI.UserDefinedFunction,
   },
 )
   return true

src/utils.jl

@@ -426,41 +426,46 @@ function parser_MOI(moimodel, index_map, nvar)
 end
 
 # Affine or quadratic, nothing to do
-function _nlp_model(model::Union{Nothing, MOI.Nonlinear.Model}, ::MOI.ModelLike, ::Type, ::Type)
-  return model
-end
+_nlp_model(::MOI.Nonlinear.Model, ::MOI.ModelLike, ::Type, ::Type) = false
 
 function _nlp_model(
-  dest::Union{Nothing, MOI.Nonlinear.Model},
+  dest::MOI.Nonlinear.Model,
   src::MOI.ModelLike,
   F::Type{<:Union{MOI.ScalarNonlinearFunction, MOI.VectorNonlinearFunction}},
   S::Type,
 )
+  has_nonlinear = false
   for ci in MOI.get(src, MOI.ListOfConstraintIndices{F, S}())
-    if isnothing(dest)
-      dest = MOI.Nonlinear.Model()
-    end
     MOI.Nonlinear.add_constraint(
       dest,
       MOI.get(src, MOI.ConstraintFunction(), ci),
       MOI.get(src, MOI.ConstraintSet(), ci),
     )
+    has_nonlinear = true
   end
-  return dest
+  return has_nonlinear
 end
 
-function _nlp_model(model::MOI.ModelLike)
-  nlp_model = nothing
+function _nlp_model(model::MOI.ModelLike)::Union{Nothing,MOI.Nonlinear.Model}
+  nlp_model = MOI.Nonlinear.Model()
+  has_nonlinear = false
+  for attr in MOI.get(model, MOI.ListOfModelAttributesSet())
+    if attr isa MOI.UserDefinedFunction
+      has_nonlinear = true
+      args = MOI.get(model, attr)
+      MOI.Nonlinear.register_operator(nlp_model, attr.name, attr.arity, args...)
+    end
+  end
   for (F, S) in MOI.get(model, MOI.ListOfConstraintTypesPresent())
-    nlp_model = _nlp_model(nlp_model, model, F, S)
+    has_nonlinear |= _nlp_model(nlp_model, model, F, S)
   end
   F = MOI.get(model, MOI.ObjectiveFunctionType())
   if F <: MOI.ScalarNonlinearFunction
-    if isnothing(nlp_model)
-      nlp_model = MOI.Nonlinear.Model()
-    end
-    attr = MOI.ObjectiveFunction{F}()
-    MOI.Nonlinear.set_objective(nlp_model, MOI.get(model, attr))
+    MOI.Nonlinear.set_objective(nlp_model, MOI.get(model, MOI.ObjectiveFunction{F}()))
+    has_nonlinear = true
+  end
+  if !has_nonlinear
+    return nothing
   end
   return nlp_model
 end

test/MOI_wrapper.jl

@@ -42,6 +42,8 @@ function test_runtests()
       r"test_nonlinear_hs071_no_hessian$",
       r"test_nonlinear_hs071_hessian_vector_product$",
       r"test_nonlinear_hs071$",
+      # Require hessian = false in the constructor of MathOptNLPModel
+      r"test_nonlinear_expression_multivariate_function$",
       # FIXME We should look over all attributes set and error
       #       for unknown ones instead of just getting the ones we know we support
       #       and ignoring the rest

test/nlp_problems/operatorspb.jl

@@ -0,0 +1,47 @@
+function operatorspb()
+  nlp = Model()
+  x0 = [390, 1000, 419.5, 340.5, 198.175, 0.5]
+  lvar = [0, 0, 340, 340, -1000, 0]
+  uvar = [400, 1000, 420, 420, 10000, 0.5236]
+  @variable(nlp, lvar[i] <= x[i = 1:6] <= uvar[i], start = x0[i])
+
+  a = 131078 // 1000
+  b = 148577 // 100000
+  ci = 90798 // 100000
+  d = cos(147588 // 100000)
+  e = sin(147588 // 100000)
+
+  @constraint(nlp, 300 - x[1] - 1 / a * x[3] * x[4] * cos(b - x[6]) + ci / a * d * x[3] == 0)
+  @constraint(nlp, -x[2] - 1 / a * x[3] * x[4] * cos(b + x[6]) + ci / a * d * x[4]^2 == 0)
+  @constraint(nlp, -x[5] - 1 / a * x[3] * x[4] * cos(b + x[6]) + ci / a * e * x[4]^2 == 0)
+  @constraint(nlp, 200 - 1 / a * x[3] * x[4] * sin(b - x[6]) + ci / a * e * x[3]^2 == 0)
+
+  function f1(t)
+    return if 0 <= t <= 300
+      30 * t
+    elseif 300 <= t <= 400
+      31 * t
+    else
+      eltype(x)(Inf)
+    end
+  end
+
+  function f2(t)
+    return if 0 <= t <= 100
+      28 * t
+    elseif 100 <= t <= 200
+      29 * t
+    elseif 200 <= t <= 1000
+      30 * t
+    else
+      eltype(t)(Inf)
+    end
+  end
+  @operator(nlp, op_f1, 1, f1)
+  @expression(nlp, op_f1)
+  @operator(nlp, op_f2, 1, f2)
+  @expression(nlp, op_f2)
+  @objective(nlp, Min, op_f1(x[1]) + op_f2(x[2]))
+
+  return nlp
+end

test/runtests.jl

@@ -5,7 +5,7 @@ using Test, Printf
 nlp_problems = setdiff(NLPModelsTest.nlp_problems, ["MGH01Feas"])
 nls_problems = NLPModelsTest.nls_problems
 
-extra_nlp_problems = ["nohesspb", "hs61", "hs100", "hs219", "quadcon"]
+extra_nlp_problems = ["nohesspb", "hs61", "hs100", "hs219", "quadcon", "operatorspb"]
 extra_nls_problems = ["nlsnohesspb", "HS30", "HS43", "MGH07", "nlsqc"]
 
 for problem in lowercase.(nlp_problems ∪ extra_nlp_problems)