Merge pull request #689 from DhairyaLGandhi/dg/obsfn

Feat: adjoints through observable functions

Author: ChrisRackauckas

.github/workflows/CI.yml

@@ -14,6 +14,7 @@ jobs:
   test:
     runs-on: ubuntu-latest
     strategy:
+      fail-fast: false
       matrix:
         group:
           - Core
@@ -47,4 +48,4 @@ jobs:
         with:
           file: lcov.info
           token: ${{ secrets.CODECOV_TOKEN }}
-          fail_ci_if_error: true
+          fail_ci_if_error: false

ext/SciMLBaseZygoteExt.jl

@@ -7,8 +7,10 @@ using SciMLBase
 using SciMLBase: ODESolution, remake,
                  getobserved, build_solution, EnsembleSolution,
                  NonlinearSolution, AbstractTimeseriesSolution
-using SymbolicIndexingInterface: symbolic_type, NotSymbolic, variable_index
+using SymbolicIndexingInterface: symbolic_type, NotSymbolic, variable_index, is_observed,
+                                 observed, parameter_values, state_values, current_time
 using RecursiveArrayTools
+import SciMLStructures
 
 # This method resolves the ambiguity with the pullback defined in
 # RecursiveArrayToolsZygoteExt
@@ -109,7 +111,18 @@ end
 @adjoint function Base.getindex(VA::ODESolution, sym)
     function ODESolution_getindex_pullback(Δ)
         i = symbolic_type(sym) != NotSymbolic() ? variable_index(VA, sym) : sym
-        if i === nothing
+        if is_observed(VA, sym)
+            f = observed(VA, sym)
+            p = parameter_values(VA)
+            tunables, _, _ = SciMLStructures.canonicalize(SciMLStructures.Tunable(), p)
+            u = state_values(VA)
+            t = current_time(VA)
+            y, back = Zygote.pullback(u, tunables) do u, tunables
+                f.(u, Ref(tunables), t)
+            end
+            gs = back(Δ)
+            (gs[1], nothing)
+        elseif i === nothing
             throw(error("Zygote AD of purely-symbolic slicing for observed quantities is not yet supported. Work around this by using `A[sym,i]` to access each element sequentially in the function being differentiated."))
         else
             Δ′ = [[i == k ? Δ[j] : zero(x[1]) for k in 1:length(x)]
@@ -120,26 +133,49 @@ end
     VA[sym], ODESolution_getindex_pullback
 end
 
+function obs_grads(VA, sym, obs_idx, Δ)
+    y, back = Zygote.pullback(VA) do sol
+        getindex.(Ref(sol), sym[obs_idx])
+    end
+    Δreduced = reduce(hcat, Δ)
+    Δobs = eachrow(Δreduced[obs_idx, :])
+    back(Δobs)
+end
+
+function obs_grads(VA, sym, ::Nothing, Δ)
+    Zygote.nt_nothing(VA)
+end
+
+function not_obs_grads(VA::ODESolution{T}, sym, not_obss_idx, i, Δ) where {T}
+    Δ′ = map(enumerate(VA.u)) do (t_idx, us)
+        map(enumerate(us)) do (u_idx, u)
+            if u_idx in i
+                idx = findfirst(isequal(u_idx), i)
+                Δ[t_idx][idx]
+            else
+                zero(T)
+            end
+        end
+    end
+
+    Δ′
+end
+
 @adjoint function Base.getindex(
         VA::ODESolution{T}, sym::Union{Tuple, AbstractVector}) where {T}
     function ODESolution_getindex_pullback(Δ)
         sym = sym isa Tuple ? collect(sym) : sym
         i = map(x -> symbolic_type(x) != NotSymbolic() ? variable_index(VA, x) : x, sym)
-        if i === nothing
-            throw(error("Zygote AD of purely-symbolic slicing for observed quantities is not yet supported. Work around this by using `A[sym,i]` to access each element sequentially in the function being differentiated."))
-        else
-            Δ′ = map(enumerate(VA.u)) do (t_idx, us)
-                map(enumerate(us)) do (u_idx, u)
-                    if u_idx in i
-                        idx = findfirst(isequal(u_idx), i)
-                        Δ[t_idx][idx]
-                    else
-                        zero(T)
-                    end
-                end
-            end
-            (Δ′, nothing)
-        end
+
+        obs_idx = findall(s -> is_observed(VA, s), sym)
+        not_obs_idx = setdiff(1:length(sym), obs_idx)
+
+        gs_obs = obs_grads(VA, sym, isempty(obs_idx) ? nothing : obs_idx, Δ)
+        gs_not_obs = not_obs_grads(VA, sym, not_obs_idx, i, Δ)
+
+        a = Zygote.accum(gs_obs[1], gs_not_obs)
+
+        (a, nothing)
     end
     VA[sym], ODESolution_getindex_pullback
 end

test/downstream/Project.toml

@@ -3,6 +3,7 @@ BoundaryValueDiffEq = "764a87c0-6b3e-53db-9096-fe964310641d"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 JumpProcesses = "ccbc3e58-028d-4f4c-8cd5-9ae44345cda5"
 ModelingToolkit = "961ee093-0014-501f-94e3-6117800e7a78"
+ModelingToolkitStandardLibrary = "16a59e39-deab-5bd0-87e4-056b12336739"
 NonlinearSolve = "8913a72c-1f9b-4ce2-8d82-65094dcecaec"
 Optimization = "7f7a1694-90dd-40f0-9382-eb1efda571ba"
 OptimizationMOI = "fd9f6733-72f4-499f-8506-86b2bdd0dea1"
@@ -22,6 +23,7 @@ BoundaryValueDiffEq = "5"
 ForwardDiff = "0.10"
 JumpProcesses = "9.10"
 ModelingToolkit = "8.37, 9"
+ModelingToolkitStandardLibrary = "2.7"
 NonlinearSolve = "2, 3"
 Optimization = "3"
 OptimizationMOI = "0.4"

test/downstream/observables_autodiff.jl

@@ -0,0 +1,102 @@
+using ModelingToolkit, OrdinaryDiffEq
+using Zygote
+using ModelingToolkit: t_nounits as t, D_nounits as D
+import SymbolicIndexingInterface as SII
+import SciMLStructures as SS
+using ModelingToolkitStandardLibrary
+import ModelingToolkitStandardLibrary as MSL
+
+@parameters σ ρ β
+@variables x(t) y(t) z(t) w(t)
+
+eqs = [D(D(x)) ~ σ * (y - x),
+    D(y) ~ x * (ρ - z) - y,
+    D(z) ~ x * y - β * z,
+    w ~ x + y + z + 2 * β]
+
+@mtkbuild sys = ODESystem(eqs, t)
+
+u0 = [D(x) => 2.0,
+    x => 1.0,
+    y => 0.0,
+    z => 0.0]
+
+p = [σ => 28.0,
+    ρ => 10.0,
+    β => 8 / 3]
+
+tspan = (0.0, 100.0)
+prob = ODEProblem(sys, u0, tspan, p, jac = true)
+sol = solve(prob, Tsit5())
+
+@testset "AutoDiff Observable Functions" begin
+    gs, = gradient(sol) do sol
+        sum(sol[sys.w])
+    end
+    du_ = [0.0, 1.0, 1.0, 1.0]
+    du = [du_ for _ in sol.u]
+    @test du == gs
+
+    # Observable in a vector
+    gs, = gradient(sol) do sol
+        sum(sum.(sol[[sys.w, sys.x]]))
+    end
+    du_ = [0.0, 1.0, 1.0, 2.0]
+    du = [du_ for _ in sol.u]
+    @test du == gs
+end
+
+# DAE
+
+function create_model(; C₁ = 3e-5, C₂ = 1e-6)
+    @variables t
+    @named resistor1 = MSL.Electrical.Resistor(R = 5.0)
+    @named resistor2 = MSL.Electrical.Resistor(R = 2.0)
+    @named capacitor1 = MSL.Electrical.Capacitor(C = C₁)
+    @named capacitor2 = MSL.Electrical.Capacitor(C = C₂)
+    @named source = MSL.Electrical.Voltage()
+    @named input_signal = MSL.Blocks.Sine(frequency = 100.0)
+    @named ground = MSL.Electrical.Ground()
+    @named ampermeter = MSL.Electrical.CurrentSensor()
+
+    eqs = [connect(input_signal.output, source.V)
+           connect(source.p, capacitor1.n, capacitor2.n)
+           connect(source.n, resistor1.p, resistor2.p, ground.g)
+           connect(resistor1.n, capacitor1.p, ampermeter.n)
+           connect(resistor2.n, capacitor2.p, ampermeter.p)]
+
+    @named circuit_model = ODESystem(eqs, t,
+        systems = [
+            resistor1, resistor2, capacitor1, capacitor2,
+            source, input_signal, ground, ampermeter
+        ])
+end
+
+@testset "DAE Observable function AD" begin
+    model = create_model()
+    sys = structural_simplify(model)
+
+    prob = ODEProblem(sys, [], (0.0, 1.0))
+    sol = solve(prob, Rodas4())
+
+    gs, = gradient(sol) do sol
+        sum(sol[sys.ampermeter.i])
+    end
+    du_ = [0.2, 1.0]
+    du = [du_ for _ in sol.u]
+    @test gs == du
+end
+
+# @testset "Adjoints with DAE" begin
+#     gs_mtkp, gs_p_new = gradient(mtkparams, p_new) do p, new_tunables
+#         new_p = SciMLStructures.replace(SciMLStructures.Tunable(), p, new_tunables)
+#         new_prob = remake(prob, p = new_p)
+#         sol = solve(new_prob, Rodas4())
+#         @show size(sol)
+#         # mean(abs.(sol[sys.ampermeter.i] .- gt))
+#         sum(sol[sys.ampermeter.i])
+#     end
+# 
+#     @test isnothing(gs_mtkp)
+#     @test length(gs_p_new) == length(p_new)
+# end

test/runtests.jl

@@ -110,6 +110,9 @@ end
         @time @safetestset "Partial Functions" begin
             include("downstream/partial_functions.jl")
         end
+        @time @safetestset "Autodiff Observable Functions" begin
+            include("downstream/observables_autodiff.jl")
+        end
     end
 
     if !is_APPVEYOR && (GROUP == "Downstream" || GROUP == "SymbolicIndexingInterface")