Why are Neural SDEs trained with TrackerAdjoint()?

I thought that Tracker was mostly replaced by Zygote and Enzyme for Reverse-mode Autodiff in the sciml ecosystem.
Why is it still being used for computing the sensitivities in DiffeqFlux?

Just limitations with higher order AD in Zygote, though those issues were actually fixed a few years ago so it could probably update its default. Also some of @avikpal 's Lux changes for higher order autodiff might come into play here too. So we probably just need to reassess it.

1 Like