These benchmarks are a work-in-progress.

Currently there is:

• small_neural_ode.py, which either evaluates, differentiates, evaluates at multiple points, or differentiates at multiple points, a small neural ODE. Try running python small_neural_ode.py with or without the --grad and --multiple flags. The results are compared against torchdiffeq and jax.experimental.ode.
• lotka_volterra.py, which solves the Lotka-Volterra equations. These results should be compared against the DifferentialEquations.jl benchmarks for the same problem (see the code).

As a general rule, Diffrax seems to match the performance of jax.experimental.ode and DifferentialEquations.jl. Meanwhile all of these are multiple times faster than torchdiffeq.