CheckParameterValue numba

[ricardoV94]

Just noticed that the numba funcify just gets rid of bound checks.

In PyMC, by default, we replace it by -inf switches here. This will obviously be slower and unnecessary for most (well) parametrized models. Just wanted to ask if it was a conscious choice.

https://github.com/aseyboldt/nutpie/blob/6f22b171ddc913e9b6b0cc16e1d25b5148db8614/nutpie/compile_pymc.py#L15

[aseyboldt]

By bounds checks you mean domain checks, right? Array accesses should still be checked, right?
But other than that, yes this should be fixed properly. I also thought there was a switch in pm.Model that disabled those domain checks, but it didn't seem to do anything.

[ricardoV94]

By bounds checks you mean domain checks, right? Array accesses should still be checked, right?
But other than that, yes this should be fixed properly. I also thought there was a switch in pm.Model that disabled those domain checks, but it didn't seem to do anything.

There is a model switch but it only disables the checks when compile_pymc (the one in aesaraf) is called. It basically switches which rewrite is included.

[ricardoV94]

It's here: https://github.com/pymc-devs/pymc/blob/a568762cd2fd0c0d19d8c0700a5e64374f1639b8/pymc/aesaraf.py#L1016

[aseyboldt]

Hm, maybe I should call that to generate the logp function.
This also reminds me that the rng code won't work properly if it ends up in the dlogp function (although I can't think of a usecase where it legitimately would end up there), because updates to shared variables won't work, and if they did they wouldn't be threadsafe.

[aseyboldt]

ref #4 for random deterministics

[aseyboldt]

nutpie is now using the pymc rewrite if bounds checks are off. (since #42)