Bugfixes for log-ODE - Githubissues

patrick-kidger commented 4 years ago

Note that I'm looking to merge this into dev-log-ode.

I noted that the diagnostics didn't run in #43, due to a RecursionError in BInterval. The reason is that the access patterns are a bad fit for what BInterval is optimised for. I've fixed this here by reordering the calls in inspect_strong_order.*
I've also increased the smallest step size in the diagnostics from 2-8 to 2-7, as I started to observe the rates misbehaving for small values, presumably because they were approaching the "analytical" discretisation.
I've also fixed a bug introduced in #43 that prevented test_strat.py from running. (The SDE was set to Ito.)
Fixed a bug that prevented log-ODE from using the default-created BrownianInterval because it didn't have the right levy_area_approximation set.

*The RecursionErrors here are a good indication that trampolining / tail recursion is probably worth adding to BInterval. I'll probably focus on other things(=running experiments for the paper) in the short term, but I'll aim to get that added in time for our next release.

lxuechen commented 4 years ago

Thanks for the PR! Though, I actually have this access order fixed in my PR locally (not pushed). Would it make sense for me to push this fix in #43 after merging #44?

patrick-kidger commented 4 years ago

Whatever you like really!

google-research / torchsde

Bugfixes for log-ODE #45