Use ADTypes.jl

.github/workflows/ci.yml

@@ -6,7 +6,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        julia_version: ["1.8", "1"]
+        julia_version: ["1.10", "1"]
     name: julia ${{ matrix.julia_version }}
     runs-on: ubuntu-latest
     steps:

Project.toml

@@ -4,6 +4,7 @@ authors = ["lassepe <[email protected]> and contributors"]
 version = "0.1.14"
 
 [deps]
+ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 FastDifferentiation = "eb9bf01b-bf85-4b60-bf87-ee5de06c00be"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
@@ -13,9 +14,10 @@ SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 Symbolics = "0c5d862f-8b57-4792-8d23-62f2024744c7"
 
 [compat]
+ADTypes = "1"
 ChainRulesCore = "1"
+FastDifferentiation = "0.3"
 ForwardDiff = "0.10"
 PATHSolver = "1.4"
 Symbolics = "4,5"
-FastDifferentiation = "0.3"
-julia = "1.8"
+julia = "1.10"

src/ParametricMCPs.jl

@@ -4,6 +4,7 @@ using PATHSolver: PATHSolver
 using SparseArrays: SparseArrays
 using FastDifferentiation: FastDifferentiation as FD
 using Symbolics: Symbolics
+using ADTypes: ADTypes
 
 include("Internals.jl")
 include("sparse_utils.jl")

src/SymbolicUtils.jl

@@ -6,10 +6,27 @@
 using Symbolics: Symbolics
 using FastDifferentiation: FastDifferentiation as FD
 
+using ADTypes: ADTypes
+
 export SymbolicsBackend, FastDifferentiationBackend, make_variables, build_function
 
-struct SymbolicsBackend end
-struct FastDifferentiationBackend end
+function SymbolicsBackend()
+    Base.depwarn(
+        "The `SymbolicsBackend` type is deprecated and will be removed in a future release. Use `ADTypes.AutoSymbolics` instead.",
+        :SymbolicUtils,
+        force = true,
+    )
+    ADTypes.AutoSymbolics()
+end
+
+function FastDifferentiationBackend()
+    Base.depwarn(
+        "The `FastDifferentiationBackend` type is deprecated and will be removed in a future release. Use `ADTypes.AutoFastDifferentiation` instead.",
+        :SymbolicUtils,
+        force = true,
+    )
+    ADTypes.AutoFastDifferentiation()
+end
 
 """
     make_variables(backend, name, dimension)
@@ -18,7 +35,7 @@
 """
 function make_variables end
 
-function make_variables(::SymbolicsBackend, name::Symbol, dimension::Int)
+function make_variables(::ADTypes.AutoSymbolics, name::Symbol, dimension::Int)
     vars = Symbolics.@variables($name[1:dimension]) |> only |> Symbolics.scalarize
 
     if isempty(vars)
@@ -28,7 +45,7 @@
     vars
 end
 
-function make_variables(::FastDifferentiationBackend, name::Symbol, dimension::Int)
+function make_variables(::ADTypes.AutoFastDifferentiation, name::Symbol, dimension::Int)
     FD.make_variables(name, dimension)
 end
 

src/parametric_problem.jl

@@ -41,7 +41,7 @@ parameter vector `θ` of size `parameter_dimension` to an length `n` vector outp
 - `parameter_dimension`: the size of the parameter vector `θ` in `f`.
 
 Keyword arguments:
-- `[backend]`: the backend (from `ParametricMCPs.SymbolicUtils`) to be used for compiling callbacks for `f` and its Jacobians needed by PATH. `SymbolicsBackend` (default) is slightly more flexible. `FastDifferentiationBackend` has reduced compilation times and reduced runtime in some cases.
+- `[backend]`: the backend (from ADTypes.jl) to be used for compiling callbacks for `f` and its Jacobians needed by PATH. `AutoSymbolics()` (default) is slightly more flexible. `AutoFastDifferentiation()` has reduced compilation times and reduced runtime in some cases.
 - `compute_sensitivities`: whether to compile the callbacks needed for sensitivity computation.
 - `[problem_size]`: the number of decision variables. If not provided and `lower_bounds` or `upper_bounds` are vectors, the problem size is inferred from the length of these vectors.
 
@@ -55,7 +55,7 @@ function ParametricMCP(
     lower_bounds,
     upper_bounds,
     parameter_dimension;
-    backend = SymbolicUtils.SymbolicsBackend(),
+    backend = ADTypes.AutoSymbolics(),
     problem_size = Internals.infer_problem_size(lower_bounds, upper_bounds),
     kwargs...,
 )

test/Project.toml

@@ -1,4 +1,5 @@
 [deps]
+ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
 FiniteDiff = "6a86dc24-6348-571c-b903-95158fe2bd41"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"

test/runtests.jl

@@ -5,9 +5,10 @@ using LinearAlgebra: norm
 using Zygote: Zygote
 using FiniteDiff: FiniteDiff
 using Symbolics: Symbolics
+using ADTypes: ADTypes
 
 @testset "ParametricMCPs.jl" begin
-    for backend in [ParametricMCPs.SymbolicUtils.SymbolicsBackend(), ParametricMCPs.SymbolicUtils.FastDifferentiationBackend()]
+    for backend in [ADTypes.AutoSymbolics(), ADTypes.AutoFastDifferentiation()]
         rng = Random.MersenneTwister(1)
         parameter_dimension = 2
         # setup a dummy MCP which represents a QP with:
@@ -16,8 +17,21 @@ using Symbolics: Symbolics
         f(z, θ) = [2z[1:2] - z[3:4] - 2θ; z[1:2]]
         lower_bounds = [-Inf, -Inf, 0, 0]
         upper_bounds = [Inf, Inf, Inf, Inf]
-        problem = ParametricMCPs.ParametricMCP(f, lower_bounds, upper_bounds, parameter_dimension; backend)
-        problem_no_jacobian = ParametricMCPs.ParametricMCP(f, lower_bounds, upper_bounds, parameter_dimension; compute_sensitivities=false, backend)
+        problem = ParametricMCPs.ParametricMCP(
+            f,
+            lower_bounds,
+            upper_bounds,
+            parameter_dimension;
+            backend,
+        )
+        problem_no_jacobian = ParametricMCPs.ParametricMCP(
+            f,
+            lower_bounds,
+            upper_bounds,
+            parameter_dimension;
+            compute_sensitivities=false,
+            backend,
+        )
 
         feasible_parameters = [[0.0, 0.0], [1.0, 0.0], [0.0, 1.0], [rand(rng, 2) for _ in 1:10]...]
         infeasible_parameters = -feasible_parameters
@@ -42,7 +56,8 @@ using Symbolics: Symbolics
 
             for θ in [feasible_parameters; infeasible_parameters]
                 ∇_autodiff_reverse = only(Zygote.gradient(dummy_pipeline, θ))
-                ∇_autodiff_forward = only(Zygote.gradient(θ -> Zygote.forwarddiff(dummy_pipeline, θ), θ))
+                ∇_autodiff_forward =
+                    only(Zygote.gradient(θ -> Zygote.forwarddiff(dummy_pipeline, θ), θ))
                 ∇_finitediff = FiniteDiff.finite_difference_gradient(dummy_pipeline, θ)
                 @test isapprox(∇_autodiff_reverse, ∇_finitediff; atol=1e-4)
                 @test isapprox(∇_autodiff_reverse, ∇_autodiff_forward; atol=1e-4)
@@ -55,7 +70,10 @@ using Symbolics: Symbolics
                 sum(solution.z .^ 2)
             end
 
-            @test_throws ArgumentError Zygote.gradient(θ -> dummy_pipeline(θ, problem_no_jacobian), feasible_parameters[1])
+            @test_throws ArgumentError Zygote.gradient(
+                θ -> dummy_pipeline(θ, problem_no_jacobian),
+                feasible_parameters[1],
+            )
         end
     end
 end