Look into adding a numba sampler

[digicosmos86]

PyMC 5.19 was just released. The version highlight is the ability to use numba to compile the likelihood function, which achieves about a 2x speed up in inference. We should look into this and see if there is an option in Bambi to allow us to use this option.

• PyMC changelog
• Bambi .fit API

[cpaniaguam]

PyMC 5.19 was just released. The version highlight is the ability to use numba to compile the likelihood function, which achieves about a 2x speed up in inference. We should look into this and see if there is an option in Bambi to allow us to use this option.

If an option for this were to be available in Bambi, wouldn't it appear in a release post-PyMC 5.19? Bambi latest release (v0.14) was published back in July.

[digicosmos86]

Just to clarify, we know that Bambi does not have this included atm. However, since bambi does provide some flexibility in using kwargs to pass additional options to PyMC's model.sample call, what we are trying to figure out is if we can take advantage of that and provide an additional option for the users to use the numba backend. Let me know if you want to chat more about this

[cpaniaguam]

I spent some time looking into this. I seems like the backends one can specify are the ones that come under bmb.inference_methods.names (see this). I couldn't find anything regarding backend options for sampling.

[AlexanderFengler]

@cpaniaguam one route to testing this is to just directly go down to the pymc model attached to any hssm model.

my_hssm_model.pymc_model, then you should be able to do something like:

import pymc as pm
with my_hssm_model.pymc_model:
    pm.sample(**kwargs)