`log_likelihood` missing in `Approximation.sample`

schmaus · September 17, 2022, 9:42am

When running this code:

with model:
    fit = pm.fit(method="advi")
trace = fit.trace()

The resulting trace inference data object is missing the log_likelihood group.

The option to create this group is explicitly set to False in this line:

github.com

pymc-devs/pymc/blob/16bee915ac42c1cf09fa1427a85aa2379b3add17/pymc/variational/opvi.py#L1458


      
              Seed for the random number generator.
          return_inferencedata : bool
              Return trace in Arviz format.
          
          
Returns
          -------
          trace: :class:`pymc.backends.base.MultiTrace`
              Samples drawn from variational posterior.
          """
          # TODO: add tests for include_transformed case
          kwargs["log_likelihood"] = False
          
          
if random_seed is not None:
              (random_seed,) = _get_seeds_per_chain(random_seed, 1)
          samples = self.sample_dict_fn(draws, random_seed=random_seed)  # type: dict
          points = ({name: records[i] for name, records in samples.items()} for i in range(draws))
          
          
trace = NDArray(
              model=self.model,
              test_point={name: records[0] for name, records in samples.items()},
          )

At first glance, setting this option to True seems to work and arviz is able to compute az.loo(trace) with a result that looks reasonable to me.

So what am I missing here? What’s the reason behind excluding log_likelihood from these samples?

ricardoV94 · September 17, 2022, 10:44am

CC @ferrine @fonnesbeck

cluhmann · September 17, 2022, 1:22pm

No idea, but there was some recent discussion of this here.

Topic		Replies	Views
"log_likelihood" not found in InferenceData Questions	5	3854	June 21, 2023
How to retrieve (log) posterior/prior/likelihood from trace? Questions	4	4097	May 23, 2020
Log-likelihood not found in InferenceData when model as pm.Potential in PYMC v5	3	247	March 27, 2024
How to evaluate Poisson log-likelihood from arviz.InferenceData after successful PyMC fitting? v5	4	601	May 7, 2023
Can't load trace in the simplest case v3 arviz	4	1202	August 18, 2022

`log_likelihood` missing in `Approximation.sample`

Related topics