Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Relax type of alg in ensemble solve for optimization #534

Merged
merged 10 commits into from
Nov 8, 2023
4 changes: 2 additions & 2 deletions src/ensemble/basic_ensemble_solve.jl
Original file line number Diff line number Diff line change
Expand Up @@ -57,10 +57,10 @@ function __solve(prob::EnsembleProblem{<:AbstractVector{<:AbstractSciMLProblem}}
end

function __solve(prob::AbstractEnsembleProblem,
alg::Union{AbstractDEAlgorithm, Nothing},
alg::A,
ensemblealg::BasicEnsembleAlgorithm;
trajectories, batch_size = trajectories,
pmap_batch_size = batch_size ÷ 100 > 0 ? batch_size ÷ 100 : 1, kwargs...)
pmap_batch_size = batch_size ÷ 100 > 0 ? batch_size ÷ 100 : 1, kwargs...) where {A}
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
pmap_batch_size = batch_size ÷ 100 > 0 ? batch_size ÷ 100 : 1, kwargs...) where {A}
pmap_batch_size = batch_size ÷ 100 > 0 ? batch_size ÷ 100 : 1, kwargs...) where {A <: SciMLAlgorithm}

I think for this to work it needs a dispatch that's not EnsembleAlgorithm?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Because of the automatic alg selection in DifferentialEquations.jl?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think I could work around this by making a stuct subtype of SciMLAlgorithm and wrap the optimizer's alg in that and handle it in __solve so we might not need to do this then

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you double check some cases with this and defaults? Tests seem to pass

num_batches = trajectories ÷ batch_size
num_batches < 1 &&
error("trajectories ÷ batch_size cannot be less than 1, got $num_batches")
Expand Down
2 changes: 2 additions & 0 deletions test/downstream/Project.toml
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
[deps]
BoundaryValueDiffEq = "764a87c0-6b3e-53db-9096-fe964310641d"
DifferentialEquations = "0c46a032-eb83-5123-abaf-570d42b7fbaa"
ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
ModelingToolkit = "961ee093-0014-501f-94e3-6117800e7a78"
Optimization = "7f7a1694-90dd-40f0-9382-eb1efda571ba"
OptimizationOptimJL = "36348300-93cb-4f02-beb5-3c3902f8871e"
Expand Down
9 changes: 9 additions & 0 deletions test/downstream/ensemble_diffeq.jl
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
using DifferentialEquations

prob = ODEProblem((u, p, t) -> 1.01u, 0.5, (0.0, 1.0))
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 isa EnsembleSolution
26 changes: 26 additions & 0 deletions test/downstream/ensemble_nondes.jl
Original file line number Diff line number Diff line change
@@ -0,0 +1,26 @@
using Optimization, OptimizationOptimJL, ForwardDiff, Test

x0 = zeros(2)
rosenbrock(x, p = nothing) = (1 - x[1])^2 + 100 * (x[2] - x[1]^2)^2
l1 = rosenbrock(x0)

optf = OptimizationFunction(rosenbrock, Optimization.AutoForwardDiff())
prob = OptimizationProblem(optf, x0)
sol1 = Optimization.solve(prob, OptimizationOptimJL.BFGS(), maxiters = 5)

ensembleprob = Optimization.EnsembleProblem(prob, [x0, x0 .+ rand(2), x0 .+ rand(2), x0 .+ rand(2)])

sol = Optimization.solve(ensembleprob, OptimizationOptimJL.BFGS(), EnsembleThreads(), trajectories = 4, maxiters = 5)
@test findmin(i -> sol[i].objective, 1:4)[1] < sol1.objective

sol = Optimization.solve(ensembleprob, OptimizationOptimJL.BFGS(), EnsembleDistributed(), trajectories = 4, maxiters = 5)
@test findmin(i -> sol[i].objective, 1:4)[1] < sol1.objective

prob = OptimizationProblem(optf, x0, lb = [-0.5, -0.5], ub = [0.5, 0.5])
ensembleprob = Optimization.EnsembleProblem(prob, 5, prob_func = (prob, i, repeat) -> remake(prob, u0 = rand(-0.5:0.001:0.5, 2)))

sol = Optimization.solve(ensembleprob, OptimizationOptimJL.BFGS(), EnsembleThreads(), trajectories = 5, maxiters = 5)
@test findmin(i -> sol[i].objective, 1:4)[1] < sol1.objective

sol = Optimization.solve(ensembleprob, OptimizationOptimJL.BFGS(), EnsembleDistributed(), trajectories = 5, maxiters = 5)
@test findmin(i -> sol[i].objective, 1:4)[1] < sol1.objective
6 changes: 6 additions & 0 deletions test/runtests.jl
Original file line number Diff line number Diff line change
Expand Up @@ -78,6 +78,12 @@ end
@time @safetestset "Ensemble solution statistics" begin
include("downstream/ensemble_stats.jl")
end
# @time @safetestset "Ensemble Optimization and Nonlinear problems" begin
# include("downstream/ensemble_nondes.jl")
# end
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this intended?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah it was to show that other tests all pass, to get this one working it needs the dispatches we have been discussing

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm lost. What's the version that works with all tests?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The one without type assertion in the __solve method

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you put it this in the form you believe works for everything and we can check the downstream tests and these tests?

@time @safetestset "Ensemble with DifferentialEquations automatic algorithm selection" begin
include("downstream/ensemble_diffeq.jl")
end
@time @safetestset "Symbol and integer based indexing of interpolated solutions" begin
include("downstream/symbol_indexing.jl")
end
Expand Down
Loading