Add only_gradient

Project.toml

@@ -1,6 +1,6 @@
 name = "AutoDiffOperators"
 uuid = "6e1301d5-4f4d-4fb5-9679-7191e22f0e0e"
-version = "0.1.6"
+version = "0.1.7"
 
 [deps]
 ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"

docs/Project.toml

@@ -3,4 +3,4 @@ Documenter = "e30172f5-a6a5-5a46-863b-614d45cd2de4"
 Markdown = "d6f4376e-aef5-505a-96c1-9c027394607a"
 
 [compat]
-Documenter = "~0.27"
+Documenter = "1"

docs/make.jl

@@ -29,7 +29,7 @@ makedocs(
     ],
     doctest = ("fixdoctests" in ARGS) ? :fix : true,
     linkcheck = !("nonstrict" in ARGS),
-    strict = !("nonstrict" in ARGS),
+    warnonly = ("nonstrict" in ARGS),
 )
 
 deploydocs(

ext/AutoDiffOperatorsForwardDiffExt.jl

@@ -52,7 +52,36 @@ end
 
 
 # ToDo: Use AD parameters
-AutoDiffOperators.with_gradient(f, x::AbstractVector{<:Real}, ad::ForwardDiffAD) = f(x), ForwardDiff.gradient(f, x)
+function AutoDiffOperators.with_gradient(f, x::AbstractVector{<:Real}, ad::ForwardDiffAD)
+    T = typeof(x)
+    U = Core.Compiler.return_type(f, Tuple{typeof(x)})
+    y = f(x)
+    R = promote_type(eltype(x), eltype(y))
+    n_y, n_x = length(y), length(x)
+    dy = similar(x, R)
+    dy .= ForwardDiff.gradient(f, x)
+    return y, dy
+end
+
+
+function AutoDiffOperators.only_gradient(f, x, ad::ForwardDiffAD)
+    T = eltype(x)
+    U = Core.Compiler.return_type(f, Tuple{typeof(x)})
+    R = promote_type(T, U)
+    _only_gradient_impl(f, x, ad, R)
+end
+
+function _only_gradient_impl(f, x, ad::ForwardDiffAD, ::Type{R}) where {R <: Real}
+    dy = similar(x, R)
+    dy .= ForwardDiff.gradient(f, x)
+    return dy
+end
+
+function _only_gradient_impl(f, x, ad::ForwardDiffAD, ::Type)
+    return ForwardDiff.gradient(f, x)
+end
+
+
 
 # ToDo: Specialize `AutoDiffOperators.with_gradient!!(f, δx, x, ad::ForwardDiffAD)`
 

src/gradient.jl

@@ -43,6 +43,19 @@ export with_gradient!!
 with_gradient!!(f, @nospecialize(δx), x, ad::ADSelector) = with_gradient(f, x, ad::ADSelector)
 
 
+"""
+    only_gradient(f, x, ad::ADSelector)
+
+Returns the gradient ∇f(x) of `f` at `x`.
+
+See also [`with_gradient(f, x, ad)`](@ref).
+"""
+function only_gradient end
+export only_gradient
+
+only_gradient(f, x, ad::ADSelector) = with_gradient(f, x, ad)[2]
+
+
 
 struct _ValGradFunc{F,AD} <: Function
     f::F
@@ -72,7 +85,7 @@ struct _GenericGradientFunc{F,AD} <: Function
 end
 _GenericGradientFunc(::Type{FT}, ad::AD) where {FT,AD<:ADSelector}  = _GenericGradientFunc{Type{FT},AD}(FT, ad)
 
-(f::_GenericGradientFunc)(x) = with_gradient(f.f, x, f.ad)[2]
+(f::_GenericGradientFunc)(x) = only_gradient(f.f, x, f.ad)
 
 """
     gradient_func(f, ad::ADSelector)

test/Project.toml

@@ -12,4 +12,4 @@ Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [compat]
-Documenter = "~0.27"
+Documenter = "1"

test/testutils.jl

@@ -39,6 +39,7 @@ function test_adsel_functionality(ad::ADSelector)
         @test_deprecated jacobian_matrix(f, x, ad) ≈ J_f_ref
         @test with_gradient(g, x, ad)[1] ≈ y_g_ref
         @test with_gradient(g, x, ad)[2] ≈ grad_g_x_ref
+        @test only_gradient(g, x, ad) ≈ grad_g_x_ref
 
         let δx = similar(x)
             fill!(δx, NaN)
@@ -57,6 +58,7 @@ function test_adsel_functionality(ad::ADSelector)
 
         if AutoDiffOperators.supports_structargs(reverse_ad_selector(ad))
             @test approx_cmp(with_gradient(f_nv, x_nv, ad), (y_nv_ref, grad_nv_ref))
+            @test approx_cmp(only_gradient(f_nv, x_nv, ad), grad_nv_ref)
             @test approx_cmp(valgrad_func(f_nv, ad)(x_nv), (y_nv_ref, grad_nv_ref))
             @test approx_cmp(gradient_func(f_nv, ad)(x_nv), grad_nv_ref)
         end