Consider changing rcParams default for log_likelihood to False

[ricardoV94]

arviz/arvizrc.template

Line 10 in 2d88638

data.log_likelihood : true # save pointwise log likelihood values, one of "true", "false"

Paraphrasing from the PyMC issue: pymc-devs/pymc#6266

This evaluation can be quite expensive and require too much memory. I think it's safe to say most users don't make use of this information making it pretty wasteful.

Would require coordinating with other PPLs that use this variable to see if they agree and can help users make the transition.

CC @OriolAbril

[ilaiyengar]

Hi, I am a first time contributor! Would this be a good first issue that I can assist with?

[ahartikainen]

Is pymc still using arviz converter?

[OriolAbril]

Hi, I am a first time contributor! Would this be a good first issue that I can assist with?

@ilaiyengar no, sorry. This issue still needs discussion and there is no clear fix yet. You should take a look at Beginner or let us know the areas within ArviZ you are interested in on gitter so we can find something that fits your interests.

Is pymc still using arviz converter?

The converter to InferenceData for pymc v4 is now in pymc codebase, but for now it maintains compatibility with ArviZ, uses it's rcParams, uses dict_to_dataset...

---

I have checked the converters and the data.log_likelihood rcParam is used by pyro, numpyro (and pymc) to compute the pointwise log likelihood values and by stan converters to retrieve log_lik variable into the log likelihood group if present. I generally use the log likelihood values but I have run into memory errors a few times after forgetting to set the value to false.

When that happens I generally need to exit the python interpreter and end up needing to resample. Therefore, given it is an rcParam that can be set again to true on a user or project basis, I think it would make sense to set the default to false, but it will probably confuse users for a while when they try to call az.compare and it doesn't work out of the box anymore.

If set to false (or whatever we choose really) it might also make sense to split the rcParam into two in order to preserve Stan behaviour where true doesn't always compute and store pointwise log likelihood and it does only if the log_lik variable is present.