Log name of RVs that are being resampled in posterior predictive

[ricardoV94]

Right now, the behavior of sample_posterior_predictive is very opaque as to which RVs are being resampled. Any RVs that are downstream of Shared variables or other variables mentioned in var_names, as well as variables missing in the trace will be resampled. We should output their names, similarly to how we show what variables are being sampled in pm.sample

For instance, in the example below it might unclear that y is being resampled, as it does not show in any log messages or in the output of posterior_predictive:

import pymc as pm
with pm.Model() as m:
    x = pm.MutableData("x", 0)
    y = pm.Normal("y", x)
    z = pm.Normal("z", y, observed=0)
    idata = pm.sample(chains=1, tune=0, draws=5, random_seed=1)
    pm.set_data({"x": 100})
    pm.sample_posterior_predictive(idata, extend_inferencedata=True, random_seed=1)
    
print(idata.posterior_predictive.data_vars)  # z (chain, draw) float64 101.8 101.5 ...
print(idata.posterior_predictive["z"].mean().values)  # 100.30893015576387

[ricardoV94]

CC @lucianopaz

[canyon289]

This does not seem very beginner friendly. Source, I'm in a room with beginners and its not entirely straightforward for them. Still a great feature, I'll remove the label

[bherwerth]

I've been trying to understand what would be needed to implement this feature. What is not yet clear to me:

1. What is the exact output of pm.sample that should be implemented analogously for pm.sample_posterior_predictive? Is it the printed message NUTS: [y]?

2. For the above example, I tried to find out where the relevant variables [y, z] explicitly occur in sample_posterior_predictive. I find vars_to_sample = [z] and vars_in_trace = [y]. However, at the point when the aesara function is created, I find z in outputs, but don't find y anymore in any of the arguments passed to aesara.function:

   pymc/pymc/aesaraf.py

   Lines 1034 to 1040 in a229983

   	aesara_function = aesara.function(
   	inputs,
   	outputs,
   	updates={**rng_updates, **kwargs.pop("updates", {})},
   	mode=mode,
   	**kwargs,
   	)

   This makes me think that perhaps sampling of y is only implicit through z.owner. Is this conclusion correct? Indeed, when I only draw from z, I also get the correct posterior predictive:

print(pm.draw(z, draws=1000).mean()) # 99.984246138739

[ricardoV94]

Yes, because y depends on a shared variable, it is resampled. But it doesn't show in the outputs.

However, if this is indeed the source of the issue #6047, we should rethink our strategy. It might be better to change our default and not resample (non-deterministic) variables that exist in the trace and emit some warning like:

"variable y which depends on a shared variable will not be resampled because it's already present in the trace. Include it in var_names if you wish to resample it.”

Although having the warning when you are okay with the behavior is annoying.

We should only be resampling deterministics with shared variables by default, not pre-existing RVs with shared variables

CC @lucianopaz

[ricardoV94]

Is it the printed message NUTS: [y]?

Yes. Something like Sampling: [y, z] in this case