ChrisRackauckas
commented
1 year ago Can you share ]st -m so we know what version of Enzyme?
Anyway to disable Enzyme in the script and just use Zygote?
Yes, just pass a sensealg and choose the autojacvec=ZygoteVJP(). For example:
Array(solve(_prob, AutoTsit5(Rodas5()), saveat = T,
abstol = 1e-6, reltol = 1e-6, verbose=false, sensealg=QuadratureAdjoint(autojacvec=ZygoteVJP())))That should be a workaround.
@wsmoses it looks like this is pointing to an issue with the GC mark loop after Enzyme? Is that something that's known? Cross referencing things like https://github.com/SciML/DiffEqFlux.jl/issues/812, it looks a bit different from the Enzyme errors I've seen before and it seems not fixed on master so I don't know if this one has been reported.
The MWE part is:
using Lux, ComponentArrays
rbf(x) = exp.(-(x .^ 2))
# Multilayer FeedForward
U = Lux.Chain( Lux.Dense(6, 25, rbf),
Lux.Dense(25, 25, rbf),
Lux.Dense(25, 25, rbf),
Lux.Dense(25, 6))
ps, st = Lux.setup(rng, U)
p_ = zeros(4)
p = ComponentArray{Float64}(p=ps, p_true=p_)
function nn_dynamics!(du, u, p, t)
û = U(u, p.p, st)[1]
du[1] = p.p_true[1] * u[1] + û[1]
du[2] = -p.p_true[4] * u[2] + û[2]
du[3] = p.p_true[1] * u[1] + û[3]
du[4] = -p.p_true[4] * u[2] + û[4]
du[5] = p.p_true[1] * u[1] + û[5]
du[6] = -p.p_true[4] * u[2] + û[6]
endand then just differentiating nn_dynamics!(du, u, p, t). Though I don't know if you'll recreate a GC issue in just one pass, while in the differential equation it'll do this derivative a few thousand times.
wsmoses
ghost
Hi,
I am having problems running a uODE in SciML. Whenever I run my script in
REPLofvscodeI do not get errors.When I run the script in the terminal with
julia ./script.jlI get a segmentation fault:Here is the script:
This is mock-example of the LV tutorial case, solving 3x the same equations using batches.
The solution I am getting when I run the script in the terminal
julia ./myscript.jlis:When I run it in
REPLofvscodeI do not see the warning:However, sometimes I also get a segmentation fault when I do
julia ./myscript.jl:Why am I getting a segmentation fault? Is the code above using memory it shouldn't? Or creating some sort of undefined behavior?
Using the same code structure, for more complex ODE's I have also seen the warning:
warning: didn't implement memmove, using memcpy as fallback which can result in errorsNothing else appeared in
REPL. I did agrepunder the.juliafolder and this warning only exists inEnzymelibrary. Anyway to disableEnzymein the script and just useZygote?Julia 1.9.0Package versions:Best Regards