Problems with bayesian ODEs parameter estimation

Questions
3

Hi,
The performance issue is somewhat diffusely related to how we implemented the DifferentialEquation with theano. There are three options:

  • one can use a different sampler, that doesn’t use gradients (wastes performance on gradients though)
  • one can use a custom Op to wrap around scipy.integrate.odeint
  • @aseyboldt is working on sunode which is 10-50x faster than what we have now. The API is not yet stable, but there’s a LotkaVoltera ODE example in the README.

I strongly believe that option 3 is the way to go. Unfortunately, I don’t have enough time these days to be of much help.

1 Like