Submitted by **martenlienen** t3_zfvb8h
in **MachineLearning**

Paper: https://arxiv.org/abs/2210.12375

Code: https://github.com/martenlienen/torchode

We have developed a new ODE solver suite for PyTorch that eliminates some unintended side-effects that can occur in batched training with adaptive step sizes by tracking a separate solver state for each sample in a batch. Additionally, torchode can speed up your neural ODE or continuous normalizing flow by minimizing the solver overhead through various implementation optimizations in its code such as combined operations (einsum, addcmul), polynomial evaluation via Horner's rule and JIT compilation. See the paper for details.

I am happy to answer questions here on reddit. If you are a NeurIPS (+workshops) attendee, it would be great to see you at my poster at the DLDE workshop on Friday at 05:10 PT / 13:10 UTC or 09:05 PT / 18:05 UTC.

MathChieft1_izehk9o wroteJudging by the Van der Pol benchmark, your methods are faster than other neural ODE solvers, but are still several magnitude slower than traditional solvers such as Runge-Kutta. Why are many people excited about this?