-
-
Notifications
You must be signed in to change notification settings - Fork 74
/
Copy pathalternative_ad_frontend.jl
285 lines (231 loc) · 12.3 KB
/
alternative_ad_frontend.jl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
using OrdinaryDiffEq, SciMLSensitivity, ForwardDiff, Zygote, ReverseDiff, Tracker, Enzyme,
FiniteDiff, Mooncake
using Test
Enzyme.API.typeWarning!(false)
mooncake_gradient(f, x) = Mooncake.value_and_gradient!!(build_rrule(f, x), f, x)[2][2]
odef(du, u, p, t) = du .= u .* p
const prob = ODEProblem(odef, [2.0], (0.0, 1.0), [3.0])
struct senseloss0{T}
sense::T
end
function (f::senseloss0)(u0p)
prob = ODEProblem{true}(odef, u0p[1:1], (0.0, 1.0), u0p[2:2])
sum(solve(prob, Tsit5(), abstol = 1e-12, reltol = 1e-12, saveat = 0.1))
end
u0p = [2.0, 3.0]
du0p = zeros(2)
dup = Zygote.gradient(senseloss0(InterpolatingAdjoint()), u0p)[1]
Enzyme.autodiff(Reverse, senseloss0(InterpolatingAdjoint()), Active, Duplicated(u0p, du0p))
dup_mc = mooncake_gradient(senseloss0(InterpolatingAdjoint()), u0p)
@test du0p ≈ dup
@test dup_mc ≈ dup
struct senseloss{T}
sense::T
end
function (f::senseloss)(u0p)
sum(solve(prob, Tsit5(), u0 = u0p[1:1], p = u0p[2:2], abstol = 1e-12,
reltol = 1e-12, saveat = 0.1, sensealg = f.sense))
end
function loss(u0p)
sum(solve(prob, Tsit5(), u0 = u0p[1:1], p = u0p[2:2], abstol = 1e-12, reltol = 1e-12,
saveat = 0.1))
end
u0p = [2.0, 3.0]
dup = Zygote.gradient(senseloss(InterpolatingAdjoint()), u0p)[1]
@test ReverseDiff.gradient(senseloss(InterpolatingAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss(ReverseDiffAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss(TrackerAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss(ForwardDiffSensitivity()), u0p) ≈ dup
@test_broken ReverseDiff.gradient(senseloss(ForwardSensitivity()), u0p) ≈ dup # broken because ForwardSensitivity not compatible with perturbing u0
@test Tracker.gradient(senseloss(InterpolatingAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss(ReverseDiffAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss(TrackerAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss(ForwardDiffSensitivity()), u0p)[1] ≈ dup
@test_broken Tracker.gradient(senseloss(ForwardSensitivity()), u0p)[1] ≈ dup # broken because ForwardSensitivity not compatible with perturbing u0
@test ForwardDiff.gradient(senseloss(InterpolatingAdjoint()), u0p) ≈ dup
@test only(Enzyme.gradient(Reverse, senseloss(InterpolatingAdjoint()), u0p)) ≈ dup
@test_throws SciMLSensitivity.EnzymeTrackedRealError only(Enzyme.gradient(
Reverse, senseloss(ReverseDiffAdjoint()), u0p))≈dup
@test_throws SciMLSensitivity.EnzymeTrackedRealError only(Enzyme.gradient(
Reverse, senseloss(TrackerAdjoint()), u0p))≈dup
@test only(Enzyme.gradient(Reverse, senseloss(ForwardDiffSensitivity()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss(ForwardSensitivity()), u0p)) ≈ dup # broken because ForwardSensitivity not compatible with perturbing u0
@test mooncake_gradient(senseloss(InterpolatingAdjoint()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss(ReverseDiffAdjoint()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss(TrackerAdjoint()), u0p) ≈ dup
@test mooncake_gradient(senseloss(ForwardDiffSensitivity()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss(ForwardSensitivity()), u0p) ≈ dup # broken because ForwardSensitivity not compatible with perturbing u0
struct senseloss2{T}
sense::T
end
prob2 = ODEProblem((du, u, p, t) -> du .= u .* p, [2.0], (0.0, 1.0), [3.0])
function (f::senseloss2)(u0p)
sum(solve(prob2, Tsit5(), u0 = u0p[1:1], p = u0p[2:2], abstol = 1e-12,
reltol = 1e-12, saveat = 0.1, sensealg = f.sense))
end
u0p = [2.0, 3.0]
dup = Zygote.gradient(senseloss2(InterpolatingAdjoint()), u0p)[1]
@test ReverseDiff.gradient(senseloss2(InterpolatingAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss2(ReverseDiffAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss2(TrackerAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss2(ForwardDiffSensitivity()), u0p) ≈ dup
@test_broken ReverseDiff.gradient(senseloss2(ForwardSensitivity()), u0p) ≈ dup # broken because ForwardSensitivity not compatible with perturbing u0
@test Tracker.gradient(senseloss2(InterpolatingAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss2(ReverseDiffAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss2(TrackerAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss2(ForwardDiffSensitivity()), u0p)[1] ≈ dup
@test_broken Tracker.gradient(senseloss2(ForwardSensitivity()), u0p)[1] ≈ dup # broken because ForwardSensitivity not compatible with perturbing u0
@test ForwardDiff.gradient(senseloss2(InterpolatingAdjoint()), u0p) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss2(InterpolatingAdjoint()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss2(ReverseDiffAdjoint()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss2(TrackerAdjoint()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss2(ForwardDiffSensitivity()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss2(ForwardSensitivity()), u0p)) ≈ dup # broken because ForwardSensitivity not compatible with perturbing u0
@test mooncake_gradient(senseloss2(InterpolatingAdjoint()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss2(ReverseDiffAdjoint()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss2(TrackerAdjoint()), u0p) ≈ dup
@test mooncake_gradient(senseloss2(ForwardDiffSensitivity()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss2(ForwardSensitivity()), u0p) ≈ dup # broken because ForwardSensitivity not compatible with perturbing u0
struct senseloss3{T}
sense::T
end
function (f::senseloss3)(u0p)
sum(solve(prob2, Tsit5(), p = u0p, abstol = 1e-12,
reltol = 1e-12, saveat = 0.1, sensealg = f.sense))
end
u0p = [3.0]
dup = Zygote.gradient(senseloss3(InterpolatingAdjoint()), u0p)[1]
@test ReverseDiff.gradient(senseloss3(InterpolatingAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss3(ReverseDiffAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss3(TrackerAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss3(ForwardDiffSensitivity()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss3(ForwardSensitivity()), u0p) ≈ dup
@test Tracker.gradient(senseloss3(InterpolatingAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss3(ReverseDiffAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss3(TrackerAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss3(ForwardDiffSensitivity()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss3(ForwardSensitivity()), u0p)[1] ≈ dup
@test ForwardDiff.gradient(senseloss3(InterpolatingAdjoint()), u0p) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss3(InterpolatingAdjoint()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss3(ReverseDiffAdjoint()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss3(TrackerAdjoint()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss3(ForwardDiffSensitivity()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss3(ForwardSensitivity()), u0p)) ≈ dup
@test mooncake_gradient(senseloss3(InterpolatingAdjoint()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss3(ReverseDiffAdjoint()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss3(TrackerAdjoint()), u0p) ≈ dup
@test mooncake_gradient(senseloss3(ForwardDiffSensitivity()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss3(ForwardSensitivity()), u0p) ≈ dup
struct senseloss4{T}
sense::T
end
function (f::senseloss4)(u0p)
sum(solve(prob, Tsit5(), p = u0p, abstol = 1e-12,
reltol = 1e-12, saveat = 0.1, sensealg = f.sense))
end
u0p = [3.0]
dup = Zygote.gradient(senseloss4(InterpolatingAdjoint()), u0p)[1]
@test ReverseDiff.gradient(senseloss4(InterpolatingAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss4(ReverseDiffAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss4(TrackerAdjoint()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss4(ForwardDiffSensitivity()), u0p) ≈ dup
@test ReverseDiff.gradient(senseloss4(ForwardSensitivity()), u0p) ≈ dup
@test Tracker.gradient(senseloss4(InterpolatingAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss4(ReverseDiffAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss4(TrackerAdjoint()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss4(ForwardDiffSensitivity()), u0p)[1] ≈ dup
@test Tracker.gradient(senseloss4(ForwardSensitivity()), u0p)[1] ≈ dup
@test ForwardDiff.gradient(senseloss4(InterpolatingAdjoint()), u0p) ≈ dup
@test only(Enzyme.gradient(Reverse, senseloss4(InterpolatingAdjoint()), u0p)) ≈ dup
@test_throws SciMLSensitivity.EnzymeTrackedRealError only(Enzyme.gradient(
Reverse, senseloss4(ReverseDiffAdjoint()), u0p))≈dup
@test_throws SciMLSensitivity.EnzymeTrackedRealError only(Enzyme.gradient(
Reverse, senseloss4(TrackerAdjoint()), u0p))≈dup
@test only(Enzyme.gradient(Reverse, senseloss4(ForwardDiffSensitivity()), u0p)) ≈ dup
@test_broken only(Enzyme.gradient(Reverse, senseloss4(ForwardSensitivity()), u0p)) ≈ dup # broken because ForwardSensitivity not compatible with perturbing u0
@test mooncake_gradient(senseloss4(InterpolatingAdjoint()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss4(ReverseDiffAdjoint()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss4(TrackerAdjoint()), u0p) ≈ dup
@test mooncake_gradient(senseloss4(ForwardDiffSensitivity()), u0p) ≈ dup
@test_broken mooncake_gradient(senseloss4(ForwardSensitivity()), u0p) ≈ dup
solvealg_test = Tsit5()
sensealg_test = InterpolatingAdjoint()
tspan = (0.0, 1.0)
u0 = rand(4, 8)
p0 = rand(16)
f_aug(u, p, t) = reshape(p, 4, 4) * u
function loss(p)
prob = ODEProblem(f_aug, u0, tspan, p; alg = solvealg_test, sensealg = sensealg_test)
sol = solve(prob)
sum(sol[:, :, end])
end
function loss2(p)
prob = ODEProblem(f_aug, u0, tspan, p)
sol = solve(prob, solvealg_test; sensealg = sensealg_test)
sum(sol[:, :, end])
end
res1 = loss(p0)
res2 = ReverseDiff.gradient(loss, p0)
res3 = loss2(p0)
res4 = ReverseDiff.gradient(loss2, p0)
@test res1≈res3 atol=1e-14
@test res2≈res4 atol=1e-14
@test_broken res2≈Enzyme.gradient(Reverse, loss, p0) atol=1e-14
@test_broken res4≈Enzyme.gradient(Reverse, loss2, p0) atol=1e-14
# I think we're just not successfully hitting the rrule here.
@test_broken res2 ≈ mooncake_gradient(loss, p0)
res4 ≈ mooncake_gradient(loss2, p0)
# Test for recursion https://discourse.julialang.org/t/diffeqsensitivity-jl-issues-with-reversediffadjoint-sensealg/88774
function ode!(derivative, state, parameters, t)
derivative .= parameters
end
function ode(state, parameters, t)
return ode!(similar(state), state, parameters, t)
end
function solve_euler(state, times, parameters)
problem = ODEProblem{true}(ode!, state, times[[1, end]], parameters; saveat = times,
sensealg = ReverseDiffAdjoint())
return solve(problem, Euler(); dt = 1e-1)
end
const initial_state = ones(2)
const solution_times = [1.0, 2.0]
ReverseDiff.gradient(p -> sum(sum(solve_euler(initial_state, solution_times, p))), zeros(2))
# Enzyme.gradient(Reverse, p -> sum(sum(solve_euler(initial_state, solution_times, p))), zeros(2))
# mooncake_gradient(p -> sum(sum(solve_euler(initial_state, solution_times, p))), zeros(2))
# https://github.com/SciML/SciMLSensitivity.jl/issues/943
GRAVITY = 9.81
MASS = 1.0
NUM_STATES = 2
t_start = 0.0
t_step = 0.05
t_stop = 2.0
tData = t_start:t_step:t_stop
u0 = [1.0, 0.0] # start state: ball position (1.0) and velocity (0.0)
p = [GRAVITY, MASS]
# setup BouncingBallODE
function fx(u, p, t)
g, m = p
return [u[2], -g]
end
ff = ODEFunction{false}(fx)
prob3 = ODEProblem{false}(ff, u0, (t_start, t_stop), p)
function loss2(p)
solution = solve(prob3; p = p, alg = solver, saveat = tData,
sensealg = sensealg, abstol = 1e-10, reltol = 1e-10)
# fix for ReverseDiff
if !isa(solution, ReverseDiff.TrackedArray) && !isa(solution, Array)
sum(abs.(collect(u[1] for u in solution.u)))
else
sum(abs.(solution[1, :]))
end
end
solver = Rosenbrock23(autodiff = false)
sensealg = ReverseDiffAdjoint()
grad_fi = FiniteDiff.finite_difference_gradient(loss2, p)
grad_fd = ForwardDiff.gradient(loss2, p)
grad_zg = Zygote.gradient(loss2, p)[1]
grad_rd = ReverseDiff.gradient(loss2, p)
@test grad_fd≈grad_fi atol=1e-2
@test grad_fd ≈ grad_zg atol=1e-4
@test grad_fd ≈ grad_rd atol=1e-4
@test_broken mooncake_gradient(loss2, p) ≈ grad_rd atol=1e-4 # appears to not be hitting the rule