End-to-end gradients with Enzyme when training ODE parameters

When reading the SciMLSensitivity docs I am somewhat surprised that Enzyme.jl is only mentioned as an option for JVPs.

Why is doing e.g. end to end gradient computations with Enzyme (e.g. much like ForwardDiff.jl does) not a more prominent option?

I suspect that this might not be very user friendly or might limit us to certain solvers due to Enzyme supporting only a subset of Julia, but I wonder if this is regardless a promising option for certain scenarios?

It only supports a subset of Julia which doesn’t include the ODE solvers at this time. But it’s growing, and that’s one our goals.

1 Like