test Flux 0.13

[ChrisRackauckas]

No description provided.

[ChrisRackauckas]

MWE of the stuff ODE solver problem:

using Flux, Zygote, ForwardDiff
import ForwardDiff: Dual

y = Float32[1.2449092, 0.26629877]
p = Float32[0.14421135, -0.006150621, 0.0393358, 0.138404, 0.23749629, 0.06463469, -0.029445898, -0.33279192, -0.094798535, -0.257304, -0.3355695, -0.1959481, 0.12938745, 0.14058144, 0.32916018, -0.23945713, -0.18813372, -0.14978944, 0.18167028, -0.22040617, -0.16580728, -0.09962158, -0.12878253, 0.24638167, -0.03310824, 0.07440266, 0.03885393, 0.27210253, -0.053823117, -0.14623246, -0.034661364, 0.049675502, 0.16398363, -0.30591217, 0.18999895, -0.26469624, 0.28702003, 0.20897748, -0.32785562, -0.100942954, -0.32169065, 0.21481845, 0.09703442, 0.30915034, 0.09057236, -0.15546058, -0.24163458, -0.13516225, -0.06676043, -0.1966813, 0.12077151, 0.056194287, -0.16526969, -0.2222915, -0.19672059, -0.034455374, -0.24578816, 0.18768719, -0.23405759, 0.046496972, -0.258523, 0.058912445, 0.042145796, -0.13487151, -0.2644665, -0.33397835, -0.29189992, 0.13996881, -0.21306355, 0.15383047, 0.15763333, -0.27050394, 0.3312636, 0.32032087, -0.24478982, -0.1096856, 0.12329024, -0.33420125, -0.1529397, 0.013263283, -0.0321317, -0.28141057, -0.058830447, -0.033951838, -0.18657157, 0.20016932, 0.1548164, 0.028861579, 0.16291597, 0.22635445, -0.090969354, 0.1766979, -0.31983075, 0.07219792, 0.23401073, -0.07207494, 0.24587327, 0.26736307, -0.23342982, 0.08657169, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -0.19381283, 0.26416284, 0.15891802, -0.12095787, 0.1411279, -0.027013905, 0.30863065, 0.122840405, 0.21558835, -0.15353091, 0.29037133, 0.15146789, 0.045571294, 0.31934103, -0.013038424, 0.19338936, -0.08277726, 0.23436436, -0.25519383, 0.19290958, -0.07854137, 0.25316665, 0.068068326, -0.25551397, 0.29491913, -0.17783256, -0.28973922, -0.102145046, -0.052218888, -0.14526269, -0.20289962, -0.22463948, 0.24003603, -0.22635874, 0.22355433, -0.10727884, -0.27763215, 0.12205175, -0.33481315, 0.04747853, 0.22055429, 0.017725615, -0.14218004, -0.27020591, 0.27612484, 0.050210662, 0.041809335, -0.032814298, -0.21339102, 0.22898024, 0.0755317, -0.23465283, -0.109813884, -0.18060842, -0.066495314, 0.22580191, 0.3323758, 0.023281226, 0.07484222, 0.28912178, -0.27487472, 0.121484526, -0.2651789, 0.19090225, -0.003508792, -0.25500044, 0.05072003, -0.07643754, 0.24113968, 0.12844749, 0.24001858, 0.2613778, -0.2603248, 0.08254892, -0.111656696, 0.23785193, 0.32324004, 0.1750177, -0.09340208, -0.12355742, -0.25986317, -0.27915004, 0.07588966, 0.25872853, 0.21791716, 0.2401611, -0.2407115, -0.23268251, -0.30390444, -0.3009561, -0.02586944, 0.16676147, -0.110212825, -0.17888871, 0.33321387, -0.32094577, -0.25499186, 0.25705588, 0.15148534, -0.28999805, 0.0, 0.0]
t = 1.5f0
λ = Dual{ForwardDiff.Tag{OrdinaryDiffEq.OrdinaryDiffEqTag,Float32},Float32,12}[Dual{ForwardDiff.Tag{OrdinaryDiffEq.OrdinaryDiffEqTag,Float32}}(0.09447026, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0), Dual{ForwardDiff.Tag{OrdinaryDiffEq.OrdinaryDiffEqTag,Float32}}(1.4116058, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0)]

model = Chain(x -> x .^ 3,
    Dense(2, 50, tanh),
    Dense(50, 2))
p, re = Flux.destructure(model)
f(u, p, t) = re(p)(u)

_dy, back = Zygote.pullback(y, p) do u, p
    vec(f(u, p, t))
end
tmp1, tmp2 = back(λ)

Found via:

using DiffEqFlux, OrdinaryDiffEq, Test

u0 = Float32[2.0; 0.0]
datasize = 30
tspan = (0.0f0, 1.5f0)

function trueODEfunc(du, u, p, t)
    true_A = [-0.1 2.0; -2.0 -0.1]
    du .= ((u .^ 3)'true_A)'
end
t = range(tspan[1], tspan[2], length=datasize)
prob = ODEProblem(trueODEfunc, u0, tspan)
ode_data = Array(solve(prob, Tsit5(), saveat=t))

model = Chain(x -> x .^ 3,
    Dense(2, 50, tanh),
    Dense(50, 2))
neuralde = NeuralODE(model, tspan, Rodas5(), saveat=t, reltol=1e-7, abstol=1e-9)

function predict_n_ode()
    neuralde(u0)
end
loss_n_ode() = sum(abs2, ode_data .- predict_n_ode())

data = Iterators.repeated((), 10)
opt = ADAM(0.1)
cb = function () #callback function to observe training
    display(loss_n_ode())
end

# Display the ODE with the initial parameter values.
cb()

neuralde = NeuralODE(model, tspan, Rodas5(), saveat=t, reltol=1e-7, abstol=1e-9)
ps = Flux.params(neuralde)
loss1 = loss_n_ode()

xx = Ref{Any}()

Flux.train!(loss_n_ode, ps, data, opt, cb=cb)

with the change:

function _vecjacobian!(dλ, y, λ, p, t, S::TS, isautojacvec::ZygoteVJP, dgrad, dy, W) where TS<:SensitivityFunction
  @unpack sensealg, f = S
  prob = getprob(S)

  isautojacvec = get_jacvec(sensealg)
  if inplace_sensitivity(S)
    if W===nothing
      _dy, back = Zygote.pullback(y, p) do u, p
        out_ = Zygote.Buffer(similar(u))
        f(out_, u, p, t)
        vec(copy(out_))
      end
    else
      _dy, back = Zygote.pullback(y, p) do u, p
        out_ = Zygote.Buffer(similar(u))
        f(out_, u, p, t, W)
        vec(copy(out_))
      end
    end
    tmp1,tmp2 = back(λ)
    dλ[:] .= vec(tmp1)
    dgrad !== nothing && tmp2 !== nothing && (dgrad[:] .= vec(tmp2))
    dy !== nothing && (dy[:] .= vec(_dy))
  else
    if W===nothing
      _dy, back = Zygote.pullback(y, p) do u, p
        vec(f(u, p, t))
      end
    else
      _dy, back = Zygote.pullback(y, p) do u, p
        vec(f(u, p, t, W))
      end
    end
    Main.xx[] = y,p,t,λ
    tmp1, tmp2 = back(λ)
    tmp1 !== nothing && (dλ[:] .= vec(tmp1))
    dy !== nothing && (dy[:] .= vec(_dy))
    dgrad !== nothing && tmp2 !== nothing && (dgrad[:] .= vec(tmp2))
  end
  return
end

ProjectManifest.zip

[ChrisRackauckas]

using ForwardDiff, Zygote, Flux
using ForwardDiff: Dual
y = Float32[0.8564646, 0.21083355]
p = Float32[-0.2548858, -0.264061, 0.06902494, -0.23288882, -0.13166176, 0.25982612, -0.26543534, -0.29349443, 0.31963557, 0.21243489, -0.2755482, -0.04317024, 0.2678376, -0.32618907, -0.11215708, -0.20082082, -0.075056225, -0.3250112, -0.20113565, -0.2580761, 0.03797583, -0.1354496, 0.18161258, 0.3180589, 0.283674, 0.05116003, -0.07082515, 0.12914972, 0.09830813, 0.29125124, 0.32423735, 0.045021717, 0.09604585, 0.007445923, 0.12431481, 0.063025564, 0.30161184, 0.23123802, 0.30304855, -0.18616274, 0.06983177, 0.13229537, 0.26679033, 0.29119095, 0.2044387, -0.1310391, 0.06418764, -0.05145624, 0.28958446, 0.08143681, -0.26594874, 0.258198, -0.16387275, 0.23627394, -0.0025739619, 0.12877232, 0.28468516, 0.14945742, -0.09824067, 0.22391124, 0.2722607, 0.034997866, 0.021131594, -0.058169674, -0.20168333, 0.3310362, 0.29977754, 0.27228144, 0.088294245, 0.17472656, 0.030819716, 0.27218765, 0.042448767, 0.25967237, 0.18181679, 0.2810931, -0.16689181, 0.17927635, 0.32586476, -0.25481033, 0.009913104, 0.20943141, -0.13506782, -0.30059853, -0.084571846, -0.31261674, 0.11608189, 0.084546946, -0.21448077, -0.19288287, -0.22511461, 0.27675447, 0.26279518, 0.061226156, -0.2828123, -0.1394083, -0.16996919, 0.2784961, -0.0039018209, -0.1362619, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.048874624, 0.11889865, 0.01040518, 0.12694769, 0.32327807, 0.13581258, 0.10043003, -0.12258695, 0.32029858, -0.05385616, -0.28262973, -0.29426816, 0.11472986, 0.014853499, -0.055616893, -0.24432188, -0.23522359, -0.07780609, 0.16605335, 0.29451388, -0.32305816, 0.03262463, -0.28862894, 0.054972157, 0.2411704, 0.31518432, 0.2221482, -0.12357236, 0.25466782, 0.03921116, -0.087710164, 0.1594814, -0.33685195, -0.13411506, 0.04239876, 0.260748, 0.15104404, 0.24697773, -0.06698533, -0.039195247, 0.29528958, -0.19330974, -0.32768622, 0.07959501, -0.11285911, -0.031941384, -0.108291335, -0.24830729, -0.08987814, -0.04234308, 0.255426, 0.3337179, 0.18690939, -0.32503495, -0.06603645, -0.17818044, 0.10007081, -0.22569874, 0.030490262, -0.014429291, 0.13864784, 0.100892544, -0.28683808, 0.05345175, -0.12727126, 0.31637886, 0.27381366, 0.026415939, 0.20263642, 0.33452004, -0.3351626, 0.0063842274, -0.26546854, -0.24439275, -0.19636214, 0.3032137, 0.13219267, 0.20853092, -0.05988348, -0.30968776, -0.1278926, 0.33035672, -0.32249796, 0.14322737, -0.29625347, -0.17458698, -0.0010983021, 0.14215776, -0.07308902, -0.19241002, 0.1702171, 0.32165667, 0.27042934, 0.068846, 0.19114906, 0.06528145, -0.31603774, 0.049985882, -0.05847536, 0.04034526, 0.0, 0.0]
t = 1.5f0
λ = ForwardDiff.Dual{ForwardDiff.Tag{Nothing,Float32},Float32,12}[Dual{ForwardDiff.Tag{Nothing,Float32}}(0.87135935, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0), Dual{ForwardDiff.Tag{Nothing,Float32}}(1.5225363, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0)]

model = Chain(x -> x .^ 3,
    Dense(2, 50, tanh),
    Dense(50, 2))

p,re = Flux.destructure(model)
f(u, p, t) = re(p)(u)
_dy, back = Zygote.pullback(y, p) do u, p
    vec(f(u, p, t))
end
tmp1, tmp2 = back(λ)

[ChrisRackauckas]

Should work once FluxML/Optimisers.jl#65

[ChrisRackauckas]

SciML/NeuralPDE.jl#508 and SciML/DeepEquilibriumNetworks.jl#44 are dependent on this.