JinraeKim
commented
2 years ago A different example (works well):
- code
using SimulationLogger using DifferentialEquations using DiffEqSensitivity using Zygote
p = 2.0 function main(p) @Loggable function dynamics!(dx, x, p, t) @log x dx .= -p*x end
if hasmethod(dynamics!, Tuple{Any, Any, Any, Any, __LOG_INDICATOR__})
# to avoid undefined error when not adding @Loggable
log_func(x, t, integrator::DiffEqBase.DEIntegrator; kwargs...) = feedback_dynamics!(zero.(x), copy(x), integrator.p, t, __LOG_INDICATOR__(); kwargs...)
x = [1, 2, 3.0]
t = 0.0
tspan = (0.0, 1.0)
prob = ODEProblem(dynamics!, x, tspan, p)
prob.p
# integrator = init(prob, Tsit5())
# @show integrator.p
# integrator.p
# @show result = log_func(x, t, integrator)
# result.x |> sum
# endend @run main(p) @show gradient(main, p)
- result
```julia
julia> include("test/log_func.jl")
gradient(main, p) = (1.0,)
(1.0,)
ChrisRackauckas
Hi, my goal is described in SciML/DiffEqFlux.jl#661, and I tried to narrow down the whole problem into a small one to determine whether it's possible to take backpropagation with integrator interface.
The following example is in this commit, branch
iss11, SimulationLogger.jl. I found thatgradientfrom Zygote.jl does not work with the integrator constructed byinit, while we can take gradient if the output is changed asprob.p(whereprob |> typeof <: ODEProblem).What would be the best workaround for this case? Or, if this is a new feature, I would like to request this new feature for flexible usage of DiffEqFlux.jl.
p = 2.0 function main(p) @Loggable function dynamics!(dx, x, p, t) @log x dx .= -p*x end
if hasmethod(dynamics!, Tuple{Any, Any, Any, Any, __LOG_INDICATOR__})
end @run main(p) @show gradient(main, p)