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.

Sign up for GitHub

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

Jump to bottom

Extend and fix AD tests #183

Merged
merged 2 commits into from Jun 11, 2020
Merged

Extend and fix AD tests #183

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e291a02
Extend and fix AD tests
devmotion Jun 11, 2020
602c5f5
Disable some tests on 32bit for now
devmotion Jun 11, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
141 changes: 83 additions & 58 deletions test/interface/ad.jl
View file
Open in desktop
Expand Up @@ -3,88 +3,113 @@ using DelayDiffEq
import FiniteDiff
import ForwardDiff

using LinearAlgebra, Test
using Test

# Hutchinson's equation
function f(du, u, h, p, t)
du[1] = p[2] * u[1] * (1 - h(p, t - p[1])[1] / p[3])
nothing
end

h(p, t) = [p[4]]
f(u, h, p, t) = p[2] * u * (1 - h(p, t - p[1]) / p[3])
h(p, t) = p[4]

@testset "Gradient" begin
function test(p)
prob = DDEProblem(f, [p[5]], h, eltype(p).((0.0, 20.0)), copy(p);
constant_lags = [p[2]])
sol = solve(prob, MethodOfSteps(Tsit5()); abstol = 1e-14, reltol = 1e-14)
diff = @. sol[1, :] - 10 * exp(-sol.t)
dot(diff, diff)
prob = DDEProblem(f, p[5], h, eltype(p).((0.0, 10.0)), copy(p);
constant_lags = (p[1],))
sol = solve(prob, MethodOfSteps(Tsit5()); abstol = 1e-14, reltol = 1e-14,
saveat = 1.0)
sum(sol)
end

# with delay length estimation
p = [1.0, 1.0, 1.0, 0.0, 1.0]
findiff = FiniteDiff.finite_difference_gradient(test, p)
fordiff = @test_logs (:warn, r"^dt <= dtmin") ForwardDiff.gradient(test, p)
@test_broken findiff ≈ fordiff

# without delay length estimation
p = [1.0, 1.0, 0.0, 1.0]
findiff2 = FiniteDiff.finite_difference_gradient(p -> test(vcat(1, p)), p)
fordiff2 = @test_logs (:warn, r"^dt <= dtmin") ForwardDiff.gradient(p -> test(vcat(1, p)), p)
@test_broken findiff2 ≈ fordiff2
# without delay length estimation
p = [1.5, 1.0, 0.5]
findiff = FiniteDiff.finite_difference_gradient(p -> test(vcat(1, p, p[end])), p)
fordiff = ForwardDiff.gradient(p -> test(vcat(1, p, p[end])), p)
@test maximum(abs.(findiff .- fordiff)) < 1e-6

# with delay length estimation and without discontinuity
p = [1.0, 1.5, 1.0, 0.5]
findiff2 = FiniteDiff.finite_difference_gradient(p -> test(vcat(p, p[end])), p)
fordiff2 = ForwardDiff.gradient(p -> test(vcat(p, p[end])), p)
@test_broken maximum(abs.(findiff2 .- fordiff2)) < 2

# with discontinuity and without delay length estimation
p = [1.5, 1.0, 0.5, 1.0]
findiff3 = FiniteDiff.finite_difference_gradient(p -> test(vcat(1, p)), p)
fordiff3 = @test_logs (:warn, r"^dt <= dtmin") ForwardDiff.gradient(p -> test(vcat(1, p)),
p)
@test_broken maximum(abs.(findiff3 .- fordiff3)) < 9

# consistency checks
@test findiff[2:end] ≈ findiff2
@test fordiff[2:end] ≈ fordiff2
@test findiff2[2:end] ≈ findiff
@test fordiff2[2:end] ≈ fordiff
@test_broken findiff3[1:(end-1)] ≈ findiff
@test_broken fordiff3[1:(end-1)] ≈ fordiff
end

@testset "Jacobian" begin
function test(p)
prob = DDEProblem(f, [p[5]], h, eltype(p).((0.0, 20.0)), copy(p);
constant_lags = [p[2]])
solve(prob, MethodOfSteps(Tsit5()); abstol = 1e-14, reltol = 1e-14).u[end]
prob = DDEProblem(f, p[5], h, eltype(p).((0.0, 10.0)), copy(p);
constant_lags = (p[1],))
sol = solve(prob, MethodOfSteps(Tsit5()); abstol = 1e-14, reltol = 1e-14,
saveat = 1.0)
sol.u
end

# with delay length estimation
p = [1.0, 1.0, 1.0, 0.0, 1.0]
findiff = FiniteDiff.finite_difference_jacobian(test, p)
fordiff = @test_logs (:warn, r"^dt <= dtmin") ForwardDiff.jacobian(test, p)
@test_broken findiff ≈ fordiff

# without delay length estimation
p = [1.0, 1.0, 0.0, 1.0]
findiff2 = FiniteDiff.finite_difference_jacobian(p -> test(vcat(1, p)), p)
fordiff2 = @test_logs (:warn, r"^dt <= dtmin") ForwardDiff.jacobian(p -> test(vcat(1, p)), p)
@test_broken findiff2 ≈ fordiff2
p = [1.5, 1.0, 0.5]
findiff = FiniteDiff.finite_difference_jacobian(p -> test(vcat(1, p, p[end])), p)
fordiff = ForwardDiff.jacobian(p -> test(vcat(1, p, p[end])), p)
@test maximum(abs.(findiff .- fordiff)) < 1e-6

# with delay length estimation and without discontinuity
p = [1.0, 1.5, 1.0, 0.5]
findiff2 = FiniteDiff.finite_difference_jacobian(p -> test(vcat(p, p[end])), p)
fordiff2 = ForwardDiff.jacobian(p -> test(vcat(p, p[end])), p)
@test_broken maximum(abs.(findiff2 .- fordiff2)) < 3

# with discontinuity and without delay length estimation
p = [1.5, 1.0, 0.5, 1.0]
findiff3 = FiniteDiff.finite_difference_jacobian(p -> test(vcat(1, p)), p)
fordiff3 = @test_logs (:warn, r"^dt <= dtmin") ForwardDiff.jacobian(p -> test(vcat(1, p)),
p)
@test_broken maximum(abs.(findiff3 .- fordiff3)) < 10

# consistency checks
@test findiff[:, 2:end] ≈ findiff2
@test fordiff[:, 2:end] ≈ fordiff2
@test findiff2[:, 2:end] ≈ findiff
@test fordiff2[:, 2:end] ≈ fordiff
@test_broken findiff3[:, 1:(end-1)] ≈ findiff
@test_broken fordiff3[:, 1:(end-1)] ≈ fordiff
end

@testset "Hessian" begin
function test(p)
prob = DDEProblem(f, [p[5]], h, eltype(p).((0.0, 20.0)), copy(p);
constant_lags = [p[2]])
sol = solve(prob, MethodOfSteps(Tsit5()); abstol = 1e-14, reltol = 1e-14)
diff = @. sol[1, :] - 10 * exp(-sol.t)
dot(diff, diff)
prob = DDEProblem(f, p[5], h, eltype(p).((0.0, 10.0)), copy(p);
constant_lags = (p[1],))
sol = solve(prob, MethodOfSteps(Tsit5()); abstol = 1e-14, reltol = 1e-14,
saveat = 1.0)
sum(sol)
end

# with delay length estimation
p = [1.0, 1.0, 1.0, 0.0, 1.0]
findiff = FiniteDiff.finite_difference_hessian(test, p)
fordiff = @test_logs (:warn, r"^dt <= dtmin") ForwardDiff.hessian(test, p)
@test_broken findiff ≈ fordiff

# without delay length estimation
p = [1.0, 1.0, 0.0, 1.0]
findiff2 = FiniteDiff.finite_difference_hessian(p -> test(vcat(1, p)), p)
fordiff2 = @test_logs (:warn, r"^dt <= dtmin") ForwardDiff.hessian(p -> test(vcat(1, p)), p)
@test_broken findiff2 ≈ fordiff2
# without delay length estimation and without discontinuity
p = [1.5, 1.0, 0.5]
findiff = FiniteDiff.finite_difference_hessian(p -> test(vcat(1, p, p[end])), p)
fordiff = ForwardDiff.hessian(p -> test(vcat(1, p, p[end])), p)
@test maximum(abs.(findiff .- fordiff)) < 1e-5

# with delay length estimation and without discontinuity
p = [1.0, 1.5, 1.0, 0.5]
findiff2 = FiniteDiff.finite_difference_hessian(p -> test(vcat(p, p[end])), p)
fordiff2 = ForwardDiff.hessian(p -> test(vcat(p, p[end])), p)
@test_broken maximum(abs.(findiff2 .- fordiff2)) < 25

# with discontinuity and without delay length estimation
p = [1.5, 1.0, 0.5, 1.0]
findiff3 = FiniteDiff.finite_difference_hessian(p -> test(vcat(1, p)), p)
fordiff3 = @test_logs (:warn, r"^dt <= dtmin") ForwardDiff.hessian(p -> test(vcat(1, p)),
p)
@test_broken maximum(abs.(findiff3 .- fordiff3)) < 6

# consistency checks
@test findiff[2:end, 2:end] ≈ findiff2
@test fordiff[2:end, 2:end] ≈ fordiff2
@test findiff2[2:end, 2:end] ≈ findiff
@test fordiff2[2:end, 2:end] ≈ fordiff
@test_broken findiff3[1:(end-1), 1:(end-1)] ≈ findiff
@test_broken fordiff3[1:(end-1), 1:(end-1)] ≈ fordiff
end