Merge pull request #512 from pepijndevos/pv/enstats

add stats to ensemble solution

Project.toml

@@ -20,6 +20,7 @@ Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
 Markdown = "d6f4376e-aef5-505a-96c1-9c027394607a"
 PrecompileTools = "aea7be01-6a6a-4083-8856-8a6e6704d82a"
 Preferences = "21216c6a-2e73-6563-6e65-726566657250"
+Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
 RecipesBase = "3cdcf5f2-1ef4-517c-9805-6587b60abb01"
 RecursiveArrayTools = "731186ca-8d62-57ce-b412-fbd966d074cd"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"

ext/SciMLBaseZygoteExt.jl

@@ -2,7 +2,8 @@ module SciMLBaseZygoteExt
 
 using Zygote: pullback
 using ZygoteRules: @adjoint
-using SciMLBase: ODESolution, issymbollike, sym_to_index, remake, getobserved
+import ZygoteRules
+using SciMLBase: EnsembleSolution, ODESolution, issymbollike, sym_to_index, remake, getobserved
 
 # This method resolves the ambiguity with the pullback defined in
 # RecursiveArrayToolsZygoteExt
@@ -55,4 +56,25 @@ end
     VA[sym, j], ODESolution_getindex_pullback
 end
 
+ZygoteRules.@adjoint function EnsembleSolution(sim, time, converged, stats)
+    out = EnsembleSolution(sim, time, converged)
+    function EnsembleSolution_adjoint(p̄::AbstractArray{T, N}) where {T, N}
+        arrarr = [[p̄[ntuple(x -> Colon(), Val(N - 2))..., j, i]
+                   for j in 1:size(p̄)[end - 1]] for i in 1:size(p̄)[end]]
+        (EnsembleSolution(arrarr, 0.0, true), nothing, nothing, nothing)
+    end
+    function EnsembleSolution_adjoint(p̄::AbstractArray{<:AbstractArray, 1})
+        (EnsembleSolution(p̄, 0.0, true), nothing, nothing, nothing)
+    end
+    function EnsembleSolution_adjoint(p̄::EnsembleSolution)
+        (p̄, nothing, nothing, nothing)
+    end
+    out, EnsembleSolution_adjoint
+end
+
+ZygoteRules.@adjoint function ZygoteRules.literal_getproperty(sim::EnsembleSolution,
+    ::Val{:u})
+    sim.u, p̄ -> (EnsembleSolution(p̄, 0.0, true),)
+end
+
 end

src/SciMLBase.jl

@@ -11,6 +11,7 @@ using LinearAlgebra
 using Statistics
 using Distributed
 using Markdown
+using Printf
 import Preferences
 
 import Logging, ArrayInterface

src/ensemble/basic_ensemble_solve.jl

@@ -23,6 +23,12 @@ $(TYPEDEF)
 """
 struct EnsembleSerial <: BasicEnsembleAlgorithm end
 
+function merge_stats(us)
+    st = Iterators.filter(!isnothing, (hasproperty(x, :stats) ? x.stats : nothing for x in us))
+    isempty(st) && return nothing
+    reduce(merge, st)
+end
+
 function __solve(prob::AbstractEnsembleProblem,
     alg::Union{AbstractDEAlgorithm, Nothing};
     kwargs...)
@@ -64,7 +70,8 @@ function __solve(prob::AbstractEnsembleProblem,
         elapsed_time = @elapsed u = solve_batch(prob, alg, ensemblealg, 1:trajectories,
             pmap_batch_size; kwargs...)
         _u = tighten_container_eltype(u)
-        return EnsembleSolution(_u, elapsed_time, true)
+        stats = merge_stats(_u)
+        return EnsembleSolution(_u, elapsed_time, true, stats)
     end
 
     converged::Bool = false
@@ -88,8 +95,8 @@ function __solve(prob::AbstractEnsembleProblem,
         end
     end
     _u = tighten_container_eltype(u)
-
-    return EnsembleSolution(_u, elapsed_time, converged)
+    stats = merge_stats(_u)
+    return EnsembleSolution(_u, elapsed_time, converged, stats)
 end
 
 function batch_func(i, prob, alg; kwargs...)

src/ensemble/ensemble_solutions.jl

@@ -29,21 +29,23 @@ struct EnsembleSolution{T, N, S} <: AbstractEnsembleSolution{T, N, S}
     u::S
     elapsedTime::Float64
     converged::Bool
+    stats
 end
-function EnsembleSolution(sim, dims::NTuple{N}, elapsedTime, converged) where {N}
-    EnsembleSolution{eltype(eltype(sim)), N, typeof(sim)}(sim, elapsedTime, converged)
+function EnsembleSolution(sim, dims::NTuple{N}, elapsedTime, converged, stats) where {N}
+    EnsembleSolution{eltype(eltype(sim)), N, typeof(sim)}(sim, elapsedTime, converged, stats)
 end
-function EnsembleSolution(sim, elapsedTime, converged)
-    EnsembleSolution(sim, (length(sim),), elapsedTime, converged)
+function EnsembleSolution(sim, elapsedTime, converged, stats=nothing)
+    EnsembleSolution(sim, (length(sim),), elapsedTime, converged, stats)
 end # Vector of some type which is not an array
 function EnsembleSolution(sim::T, elapsedTime,
-    converged) where {T <: AbstractVector{T2}
+    converged, stats=nothing) where {T <: AbstractVector{T2}
 } where {T2 <:
          AbstractArray}
-    EnsembleSolution{eltype(eltype(sim)), ndims(sim[1]) + 1,
-        typeof(sim)}(sim,
+    EnsembleSolution{eltype(eltype(sim)), ndims(sim[1]) + 1, typeof(sim)}(
+        sim,
         elapsedTime,
-        converged)
+        converged,
+        stats)
 end
 
 struct WeightedEnsembleSolution{T1 <: AbstractEnsembleSolution, T2 <: Number}
@@ -56,7 +58,7 @@ struct WeightedEnsembleSolution{T1 <: AbstractEnsembleSolution, T2 <: Number}
 end
 
 function Base.reverse(sim::EnsembleSolution)
-    EnsembleSolution(reverse(sim.u), sim.elapsedTime, sim.converged)
+    EnsembleSolution(reverse(sim.u), sim.elapsedTime, sim.converged, sim.stats)
 end
 
 """

src/solutions/nonlinear_solutions.jl

@@ -17,6 +17,20 @@ mutable struct NLStats
     nsteps::Int
 end
 
+function Base.show(io::IO, ::MIME"text/plain", s::NLStats)
+    println(io, summary(s))
+    @printf io "%-50s %-d\n" "Number of function evaluations:" s.nf
+    @printf io "%-50s %-d\n" "Number of Jacobians created:" s.njacs
+    @printf io "%-50s %-d\n" "Number of factorizations:" s.nfactors
+    @printf io "%-50s %-d\n" "Number of linear solves:" s.nsolve
+    @printf io "%-50s %-d" "Number of nonlinear solver iterations:" s.nsteps
+end
+
+function Base.merge(s1::NLStats, s2::NLStats)
+    NLStats(s1.nf + s2.nf, s1.njacs + s2.njacs, s1.nfactors + s2.nfactors,
+        s1.nsolve + s2.nsolve, s1.nsteps + s2.nsteps)
+end
+
 """
 $(TYPEDEF)
 

src/solutions/ode_solutions.jl

@@ -1,6 +1,79 @@
 """
 $(TYPEDEF)
 
+Statistics from the differential equation solver about the solution process.
+
+## Fields
+
+- nf: Number of function evaluations. If the differential equation is a split function,
+  such as a `SplitFunction` for implicit-explicit (IMEX) integration, then `nf` is the
+  number of function evaluations for the first function (the implicit function)
+- nf2: If the differential equation is a split function, such as a `SplitFunction`
+  for implicit-explicit (IMEX) integration, then `nf2` is the number of function
+  evaluations for the second function, i.e. the function treated explicitly. Otherwise
+  it is zero.
+- nw: The number of W=I-gamma*J (or W=I/gamma-J) matrices constructed during the solving
+  process.
+- nsolve: The number of linear solves `W\b` required for the integration.
+- njacs: Number of Jacobians calculated during the integration.
+- nnonliniter: Total number of iterations for the nonlinear solvers.
+- nnonlinconvfail: Number of nonlinear solver convergence failures.
+- ncondition: Number of calls to the condition function for callbacks.
+- naccept: Number of accepted steps.
+- nreject: Number of rejected steps.
+- maxeig: Maximum eigenvalue over the solution. This is only computed if the
+  method is an auto-switching algorithm.
+"""
+mutable struct DEStats
+    nf::Int
+    nf2::Int
+    nw::Int
+    nsolve::Int
+    njacs::Int
+    nnonliniter::Int
+    nnonlinconvfail::Int
+    ncondition::Int
+    naccept::Int
+    nreject::Int
+    maxeig::Float64
+end
+
+DEStats(x::Int = -1) = DEStats(x, x, x, x, x, x, x, x, x, x, 0.0)
+
+function Base.show(io::IO, ::MIME"text/plain", s::DEStats)
+    println(io, summary(s))
+    @printf io "%-50s %-d\n" "Number of function 1 evaluations:" s.nf
+    @printf io "%-50s %-d\n" "Number of function 2 evaluations:" s.nf2
+    @printf io "%-50s %-d\n" "Number of W matrix evaluations:" s.nw
+    @printf io "%-50s %-d\n" "Number of linear solves:" s.nsolve
+    @printf io "%-50s %-d\n" "Number of Jacobians created:" s.njacs
+    @printf io "%-50s %-d\n" "Number of nonlinear solver iterations:" s.nnonliniter
+    @printf io "%-50s %-d\n" "Number of nonlinear solver convergence failures:" s.nnonlinconvfail
+    @printf io "%-50s %-d\n" "Number of rootfind condition calls:" s.ncondition
+    @printf io "%-50s %-d\n" "Number of accepted steps:" s.naccept
+    @printf io "%-50s %-d" "Number of rejected steps:" s.nreject
+    iszero(s.maxeig) || @printf io "\n%-50s %-d" "Maximum eigenvalue recorded:" s.maxeig
+end
+
+function Base.merge(a::DEStats, b::DEStats)
+    DEStats(
+        a.nf + b.nf,
+        a.nf2 + b.nf2,
+        a.nw + b.nw,
+        a.nsolve + b.nsolve,
+        a.njacs + b.njacs,
+        a.nnonliniter + b.nnonliniter,
+        a.nnonlinconvfail + b.nnonlinconvfail,
+        a.ncondition + b.ncondition,
+        a.naccept + b.naccept,
+        a.nreject + b.nreject,
+        max(a.maxeig, b.maxeig),
+    ) 
+end
+
+"""
+$(TYPEDEF)
+
 Representation of the solution to an ordinary differential equation defined by an ODEProblem.
 
 ## DESolution Interface

test/downstream/ensemble_stats.jl

@@ -0,0 +1,13 @@
+using OrdinaryDiffEq
+using Test
+
+f(u,p,t) = 1.01*u
+u0=1/2
+tspan = (0.0,1.0)
+prob = ODEProblem(f,u0,tspan)
+function prob_func(prob, i, repeat)
+    remake(prob, u0 = rand() * prob.u0)
+end
+ensemble_prob = EnsembleProblem(prob, prob_func = prob_func)
+sim = solve(ensemble_prob, Tsit5(), EnsembleThreads(), trajectories = 10)
+@test sim.stats.nf == mapreduce(x -> x.stats.nf, +, sim.u)

test/runtests.jl

@@ -75,6 +75,9 @@ end
         @time @safetestset "solving Ensembles with multiple problems" begin
             include("downstream/ensemble_multi_prob.jl")
         end
+        @time @safetestset "Ensemble solution statistics" begin
+            include("downstream/ensemble_stats.jl")
+        end
         @time @safetestset "Symbol and integer based indexing of interpolated solutions" begin
             include("downstream/symbol_indexing.jl")
         end