Fabian Fröhlich
Fabian Fröhlich
requires https://github.com/ICB-DCM/fiddy/pull/18 and https://github.com/ICB-DCM/fiddy/pull/17
> The current status is that the test for each benchmark model needs to be made to pass, e.g. by finding a vector for each model that reliably passes the...
https://github.com/wesselb/fdm might be relevant here
fantastic thanks, will keep you updated.
Okay this was quite helpful in getting to the bottom of this. The issues with breakpoints I am experiencing do appear to be related to https://github.com/google/jax/issues/16732, as setting the number...
Indeed, passing a small value (e.g. 1e-10) as `dt0` fixes the issue, at least for some values of `k`.
Okay, maybe I should have written backward _pass_. I am not familiar with the "discretise-then-optimise" implementation, but to me it's everything suggests that the back propagation fails because NaNs resulting...
> So the default discretise-then-optimise (`adjoint=RecursiveCheckpointAdjoint()`) is just the naive backpropagation through the internals of the solver. This means that we never actually evaluate the vector field mutiple times backward-in-time....
> That's great to have a standalone example, than you! I do notice that it still includes some additional complexities e.g. a `jax.vmap`. Can this example be simplified further? Yes,...
Hi, thanks for your detailed reply! I am already using `jax.config.update("jax_enable_x64", True)`, but didn't include it here since it makes it easier finding problematic parameter values. Using `expm1` is a...