Merge pull request #315 from SouthEndMusic/cache_parameters_opt_in

Refactor parameter caching + add Zygote tests

Project.toml

@@ -7,7 +7,6 @@ FindFirstFunctions = "64ca27bc-2ba2-4a57-88aa-44e436879224"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 PrettyTables = "08abe8d2-0d0c-5749-adfa-8a2ac140af0d"
-ReadOnlyArrays = "988b38a3-91fc-5605-94a2-ee2116b3bd83"
 RecipesBase = "3cdcf5f2-1ef4-517c-9805-6587b60abb01"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
 
@@ -16,6 +15,7 @@ ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 Optim = "429524aa-4258-5aef-a3af-852621145aeb"
 RegularizationTools = "29dad682-9a27-4bc3-9c72-016788665182"
 Symbolics = "0c5d862f-8b57-4792-8d23-62f2024744c7"
+Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [extensions]
 DataInterpolationsChainRulesCoreExt = "ChainRulesCore"
@@ -33,14 +33,14 @@ LinearAlgebra = "1.10"
 Optim = "1.6"
 PrettyTables = "2"
 QuadGK = "2.9.1"
-ReadOnlyArrays = "0.2.0"
 RecipesBase = "1.3"
 Reexport = "1"
 RegularizationTools = "0.6"
 SafeTestsets = "0.1"
 StableRNGs = "1"
 Symbolics = "5.29"
 Test = "1"
+Zygote = "0.6"
 julia = "1.10"
 
 [extras]
@@ -55,6 +55,7 @@ SafeTestsets = "1bc83da4-3b8d-516f-aca4-4fe02f6d838f"
 StableRNGs = "860ef19b-820b-49d6-a774-d7a799459cd3"
 Symbolics = "0c5d862f-8b57-4792-8d23-62f2024744c7"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [targets]
-test = ["Aqua", "SafeTestsets", "ChainRulesCore", "Optim", "RegularizationTools", "Test", "StableRNGs", "FiniteDifferences", "QuadGK", "ForwardDiff", "Symbolics"]
+test = ["Aqua", "SafeTestsets", "ChainRulesCore", "Optim", "RegularizationTools", "Test", "StableRNGs", "FiniteDifferences", "QuadGK", "ForwardDiff", "Symbolics", "Zygote"]

docs/src/interface.md

@@ -35,22 +35,7 @@ A2(300.0)
 
     The values computed beyond the range of the time points provided during interpolation will not be reliable, as these methods only perform well within the range and the first/last piece polynomial fit is extrapolated on either side which might not reflect the true nature of the data.
 
-The keyword `safetycopy = false` can be passed to make sure no copies of `u` and `t` are made when initializing the interpolation object.
-
-```@example interface
-A3 = QuadraticInterpolation(u, t; safetycopy = false)
-
-# Check for same memory
-u === A3.u.parent
-```
-
-Note that this does not prevent allocation in every interpolation constructor call, because parameter values are cached for all interpolation types except [`ConstantInterpolation`](@ref).
-
-Because of the caching of parameters which depend on `u` and `t`, this data should not be mutated. Therefore `u` and `t` are wrapped in a `ReadOnlyArray` from [ReadOnlyArrays.jl](https://github.com/JuliaArrays/ReadOnlyArrays.jl).
-
-```@repl interface
-A3.t[2] = 3.14
-```
+The keyword `cache_parameters = true` can be passed to precalculate parameters at initialization, making evalations cheaper to compute. This is not compatible with modifying `u` and `t`. The default `cache_parameters = false` does however not prevent allocation in every interpolation constructor call.
 
 ## Derivatives
 

ext/DataInterpolationsChainRulesCoreExt.jl

@@ -1,27 +1,97 @@
 module DataInterpolationsChainRulesCoreExt
-
 if isdefined(Base, :get_extension)
     using DataInterpolations: _interpolate, derivative, AbstractInterpolation,
+                              LinearInterpolation, QuadraticInterpolation,
                               LagrangeInterpolation, AkimaInterpolation,
-                              BSplineInterpolation, BSplineApprox
+                              BSplineInterpolation, BSplineApprox, get_idx, get_parameters,
+                              _quad_interp_indices
     using ChainRulesCore
 else
     using ..DataInterpolations: _interpolate, derivative, AbstractInterpolation,
+                                LinearInterpolation, QuadraticInterpolation,
                                 LagrangeInterpolation, AkimaInterpolation,
-                                BSplineInterpolation, BSplineApprox
+                                BSplineInterpolation, BSplineApprox, get_parameters,
+                                _quad_interp_indices
     using ..ChainRulesCore
 end
 
+function ChainRulesCore.rrule(
+        ::Type{LinearInterpolation}, u, t, I, p, extrapolate, cache_parameters)
+    A = LinearInterpolation(u, t, I, p, extrapolate, cache_parameters)
+    function LinearInterpolation_pullback(ΔA)
+        df = NoTangent()
+        du = ΔA.u
+        dt = NoTangent()
+        dI = NoTangent()
+        dp = NoTangent()
+        dextrapolate = NoTangent()
+        dcache_parameters = NoTangent()
+        df, du, dt, dI, dp, dextrapolate, dcache_parameters
+    end
+
+    A, LinearInterpolation_pullback
+end
+
+function ChainRulesCore.rrule(
+        ::Type{QuadraticInterpolation}, u, t, I, p, mode, extrapolate, cache_parameters)
+    A = QuadraticInterpolation(u, t, I, p, mode, extrapolate, cache_parameters)
+    function LinearInterpolation_pullback(ΔA)
+        df = NoTangent()
+        du = ΔA.u
+        dt = NoTangent()
+        dI = NoTangent()
+        dp = NoTangent()
+        dmode = NoTangent()
+        dextrapolate = NoTangent()
+        dcache_parameters = NoTangent()
+        df, du, dt, dI, dp, dmode, dextrapolate, dcache_parameters
+    end
+
+    A, LinearInterpolation_pullback
+end
+
+function u_tangent(A::LinearInterpolation, t, Δ)
+    out = zero(A.u)
+    idx = get_idx(A.t, t, A.idx_prev[])
+    t_factor = (t - A.t[idx]) / (A.t[idx + 1] - A.t[idx])
+    out[idx] = Δ * (one(eltype(out)) - t_factor)
+    out[idx + 1] = Δ * t_factor
+    out
+end
+
+function u_tangent(A::QuadraticInterpolation, t, Δ)
+    out = zero(A.u)
+    i₀, i₁, i₂ = _quad_interp_indices(A, t, A.idx_prev[])
+    t₀ = A.t[i₀]
+    t₁ = A.t[i₁]
+    t₂ = A.t[i₂]
+    Δt₀ = t₁ - t₀
+    Δt₁ = t₂ - t₁
+    Δt₂ = t₂ - t₀
+    out[i₀] = Δ * (t - A.t[i₁]) * (t - A.t[i₂]) / (Δt₀ * Δt₂)
+    out[i₁] = -Δ * (t - A.t[i₀]) * (t - A.t[i₂]) / (Δt₀ * Δt₁)
+    out[i₂] = Δ * (t - A.t[i₀]) * (t - A.t[i₁]) / (Δt₂ * Δt₁)
+    out
+end
+
+function u_tangent(A, t, Δ)
+    NoTangent()
+end
+
 function ChainRulesCore.rrule(::typeof(_interpolate),
         A::Union{
+            LinearInterpolation,
+            QuadraticInterpolation,
             LagrangeInterpolation,
             AkimaInterpolation,
             BSplineInterpolation,
             BSplineApprox
         },
         t::Number)
     deriv = derivative(A, t)
-    interpolate_pullback(Δ) = (NoTangent(), NoTangent(), deriv * Δ)
+    function interpolate_pullback(Δ)
+        (NoTangent(), Tangent{typeof(A)}(; u = u_tangent(A, t, Δ)), deriv * Δ)
+    end
     return _interpolate(A, t), interpolate_pullback
 end
 

ext/DataInterpolationsOptimExt.jl

@@ -18,9 +18,8 @@ function Curvefit(u,
         box = false,
         lb = nothing,
         ub = nothing;
-        extrapolate = false,
-        safetycopy = false)
-    u, t = munge_data(u, t, safetycopy)
+        extrapolate = false)
+    u, t = munge_data(u, t)
     errfun(t, u, p) = sum(abs2.(u .- model(t, p)))
     if box == false
         mfit = optimize(p -> errfun(t, u, p), p0, alg)

ext/DataInterpolationsRegularizationToolsExt.jl

@@ -69,8 +69,8 @@ A = RegularizationSmooth(u, t, t̂, wls, wr, d; λ = 1.0, alg = :gcv_svd)
 """
 function RegularizationSmooth(u::AbstractVector, t::AbstractVector, t̂::AbstractVector,
         wls::AbstractVector, wr::AbstractVector, d::Int = 2;
-        λ::Real = 1.0, alg::Symbol = :gcv_svd, extrapolate::Bool = false, safetycopy::Bool = true)
-    u, t = munge_data(u, t, safetycopy)
+        λ::Real = 1.0, alg::Symbol = :gcv_svd, extrapolate::Bool = false)
+    u, t = munge_data(u, t)
     M = _mapping_matrix(t̂, t)
     Wls½ = LA.diagm(sqrt.(wls))
     Wr½ = LA.diagm(sqrt.(wr))
@@ -86,8 +86,8 @@ A = RegularizationSmooth(u, t, d; λ = 1.0, alg = :gcv_svd, extrapolate = false)
 """
 function RegularizationSmooth(u::AbstractVector, t::AbstractVector, d::Int = 2;
         λ::Real = 1.0,
-        alg::Symbol = :gcv_svd, extrapolate::Bool = false, safetycopy::Bool = true)
-    u, t = munge_data(u, t, safetycopy)
+        alg::Symbol = :gcv_svd, extrapolate::Bool = false)
+    u, t = munge_data(u, t)
     t̂ = t
     N = length(t)
     M = Array{Float64}(LA.I, N, N)
@@ -115,8 +115,8 @@ A = RegularizationSmooth(u, t, t̂, d; λ = 1.0, alg = :gcv_svd, extrapolate = f
 """
 function RegularizationSmooth(u::AbstractVector, t::AbstractVector, t̂::AbstractVector,
         d::Int = 2; λ::Real = 1.0, alg::Symbol = :gcv_svd,
-        extrapolate::Bool = false, safetycopy::Bool = true)
-    u, t = munge_data(u, t, safetycopy)
+        extrapolate::Bool = false)
+    u, t = munge_data(u, t)
     N, N̂ = length(t), length(t̂)
     M = _mapping_matrix(t̂, t)
     Wls½ = Array{Float64}(LA.I, N, N)
@@ -143,8 +143,8 @@ A = RegularizationSmooth(u, t, t̂, wls, d; λ = 1.0, alg = :gcv_svd, extrapolat
 """
 function RegularizationSmooth(u::AbstractVector, t::AbstractVector, t̂::AbstractVector,
         wls::AbstractVector, d::Int = 2; λ::Real = 1.0,
-        alg::Symbol = :gcv_svd, extrapolate::Bool = false, safetycopy::Bool = true)
-    u, t = munge_data(u, t, safetycopy)
+        alg::Symbol = :gcv_svd, extrapolate::Bool = false)
+    u, t = munge_data(u, t)
     N, N̂ = length(t), length(t̂)
     M = _mapping_matrix(t̂, t)
     Wls½ = LA.diagm(sqrt.(wls))
@@ -172,8 +172,8 @@ A = RegularizationSmooth(
 """
 function RegularizationSmooth(u::AbstractVector, t::AbstractVector, t̂::Nothing,
         wls::AbstractVector, d::Int = 2; λ::Real = 1.0,
-        alg::Symbol = :gcv_svd, extrapolate::Bool = false, safetycopy::Bool = true)
-    u, t = munge_data(u, t, safetycopy)
+        alg::Symbol = :gcv_svd, extrapolate::Bool = false)
+    u, t = munge_data(u, t)
     t̂ = t
     N = length(t)
     M = Array{Float64}(LA.I, N, N)
@@ -202,8 +202,8 @@ A = RegularizationSmooth(
 """
 function RegularizationSmooth(u::AbstractVector, t::AbstractVector, t̂::Nothing,
         wls::AbstractVector, wr::AbstractVector, d::Int = 2;
-        λ::Real = 1.0, alg::Symbol = :gcv_svd, extrapolate::Bool = false, safetycopy::Bool = true)
-    u, t = munge_data(u, t, safetycopy)
+        λ::Real = 1.0, alg::Symbol = :gcv_svd, extrapolate::Bool = false)
+    u, t = munge_data(u, t)
     t̂ = t
     N = length(t)
     M = Array{Float64}(LA.I, N, N)
@@ -232,8 +232,8 @@ A = RegularizationSmooth(
 """
 function RegularizationSmooth(u::AbstractVector, t::AbstractVector, t̂::Nothing,
         wls::Symbol, d::Int = 2; λ::Real = 1.0, alg::Symbol = :gcv_svd,
-        extrapolate::Bool = false, safetycopy::Bool = true)
-    u, t = munge_data(u, t, safetycopy)
+        extrapolate::Bool = false)
+    u, t = munge_data(u, t)
     t̂ = t
     N = length(t)
     M = Array{Float64}(LA.I, N, N)

src/DataInterpolations.jl

@@ -7,7 +7,6 @@ abstract type AbstractInterpolation{T} end
 using LinearAlgebra, RecipesBase
 using PrettyTables
 using ForwardDiff
-using ReadOnlyArrays
 import FindFirstFunctions: searchsortedfirstcorrelated, searchsortedlastcorrelated,
                            bracketstrictlymontonic
 
@@ -90,12 +89,6 @@ function Base.showerror(io::IO, e::IntegralNotInvertibleError)
     print(io, INTEGRAL_NOT_INVERTIBLE_ERROR)
 end
 
-const MUST_COPY_ERROR = "A copy must be made of u, t to filter missing data"
-struct MustCopyError <: Exception end
-function Base.showerror(io::IO, e::MustCopyError)
-    print(io, MUST_COPY_ERROR)
-end
-
 export LinearInterpolation, QuadraticInterpolation, LagrangeInterpolation,
        AkimaInterpolation, ConstantInterpolation, QuadraticSpline, CubicSpline,
        BSplineInterpolation, BSplineApprox, CubicHermiteSpline, PCHIPInterpolation,
@@ -128,12 +121,12 @@ struct RegularizationSmooth{uType, tType, T, T2, ITP <: AbstractInterpolation{T}
             Aitp,
             extrapolate)
         new{typeof(u), typeof(t), eltype(u), typeof(λ), typeof(Aitp)}(
-            readonly_wrap(u),
-            readonly_wrap(û),
-            readonly_wrap(t),
-            readonly_wrap(t̂),
-            readonly_wrap(oftype(u.parent, wls)),
-            readonly_wrap(oftype(u.parent, wr)),
+            u,
+            û,
+            t,
+            t̂,
+            wls,
+            wr,
             d,
             λ,
             alg,

src/derivatives.jl

@@ -19,14 +19,16 @@ end
 
 function _derivative(A::LinearInterpolation, t::Number, iguess)
     idx = get_idx(A.t, t, iguess; idx_shift = -1, ub_shift = -1, side = :first)
-    A.p.slope[idx], idx
+    slope = get_parameters(A, idx)
+    slope, idx
 end
 
 function _derivative(A::QuadraticInterpolation, t::Number, iguess)
     i₀, i₁, i₂ = _quad_interp_indices(A, t, iguess)
-    du₀ = A.p.l₀[i₀] * (2t - A.t[i₁] - A.t[i₂])
-    du₁ = A.p.l₁[i₀] * (2t - A.t[i₀] - A.t[i₂])
-    du₂ = A.p.l₂[i₀] * (2t - A.t[i₀] - A.t[i₁])
+    l₀, l₁, l₂ = get_parameters(A, i₀)
+    du₀ = l₀ * (2t - A.t[i₁] - A.t[i₂])
+    du₁ = l₁ * (2t - A.t[i₀] - A.t[i₂])
+    du₂ = l₂ * (2t - A.t[i₀] - A.t[i₁])
     return @views @. du₀ + du₁ + du₂, i₀
 end
 
@@ -129,7 +131,7 @@ end
 # QuadraticSpline Interpolation
 function _derivative(A::QuadraticSpline{<:AbstractVector}, t::Number, iguess)
     idx = get_idx(A.t, t, iguess; lb = 2, ub_shift = 0, side = :first)
-    σ = A.p.σ[idx - 1]
+    σ = get_parameters(A, idx - 1)
     A.z[idx - 1] + 2σ * (t - A.t[idx - 1]), idx
 end
 
@@ -139,8 +141,9 @@ function _derivative(A::CubicSpline{<:AbstractVector}, t::Number, iguess)
     Δt₁ = t - A.t[idx]
     Δt₂ = A.t[idx + 1] - t
     dI = (-A.z[idx] * Δt₂^2 + A.z[idx + 1] * Δt₁^2) / (2A.h[idx + 1])
-    dC = A.p.c₁[idx]
-    dD = -A.p.c₂[idx]
+    c₁, c₂ = get_parameters(A, idx)
+    dC = c₁
+    dD = -c₂
     dI + dC + dD, idx
 end
 
@@ -193,7 +196,8 @@ function _derivative(
     Δt₀ = t - A.t[idx]
     Δt₁ = t - A.t[idx + 1]
     out = A.du[idx]
-    out += Δt₀ * (Δt₀ * A.p.c₂[idx] + 2(A.p.c₁[idx] + Δt₁ * A.p.c₂[idx]))
+    c₁, c₂ = get_parameters(A, idx)
+    out += Δt₀ * (Δt₀ * c₂ + 2(c₁ + Δt₁ * c₂))
     out, idx
 end
 
@@ -204,7 +208,8 @@ function _derivative(
     Δt₀ = t - A.t[idx]
     Δt₁ = t - A.t[idx + 1]
     out = A.du[idx] + A.ddu[idx] * Δt₀
+    c₁, c₂, c₃ = get_parameters(A, idx)
     out += Δt₀^2 *
-           (3A.p.c₁[idx] + (3Δt₁ + Δt₀) * A.p.c₂[idx] + (3Δt₁^2 + Δt₀ * 2Δt₁) * A.p.c₃[idx])
+           (3c₁ + (3Δt₁ + Δt₀) * c₂ + (3Δt₁^2 + Δt₀ * 2Δt₁) * c₃)
     out, idx
 end