UserWarning if doing predictive sampling with models containing Potentials

[ricardoV94]

As discussed in #3865

Do we want tests for the Warning?

[michaelosthege]

Looks good! A simple test would be great.

[AlexAndorra]

Nice, thanks @ricardoV94 ! I agree with @michaelosthege that a simple test would be nice.
Also, this should probably be added to fast_sample_posterior_predictive too: https://github.com/pymc-devs/pymc3/blob/1769258e459e8f40aa8a56e0ac911aa99e7f67de/pymc3/distributions/posterior_predictive.py#L156

[ricardoV94]

Nice, thanks @ricardoV94 ! I agree with @michaelosthege that a simple test would be nice.
Also, this should probably be added to fast_sample_posterior_predictive too:

https://github.com/pymc-devs/pymc3/blob/1769258e459e8f40aa8a56e0ac911aa99e7f67de/pymc3/distributions/posterior_predictive.py#L156

Nice catch!

I added some tests, do you think they are an overkill?

[michaelosthege]

The tests are overkill in the sense that they will add unnecessarily lots of runtime.
Instead of calling pm.sample() which will be a complete overkill in terms of compute, it would be better to use az.InferenceData.from_dict and manually create a "trace" using numpy.random.

[ricardoV94]

The tests are overkill in the sense that they will add unnecessarily lots of runtime.
Instead of calling pm.sample() which will be a complete overkill in terms of compute, it would be better to use az.InferenceData.from_dict and manually create a "trace" using numpy.random.

Agreed! By any chance, do you know if there is any test already doing something like this?

[michaelosthege]

Agreed! By any chance, do you know if there is any test already doing something like this?

It doesn't looks like that's the case so far. Many tests probably could.

You can inspect the existing trace to find out which variables you'll need. Something along the lines of

for vname in trace.varnames:
    print(f"vname is shape {trace[vname].shape}")
    print(trace[vname])

[AlexAndorra]

I guess you can also take some cues from ArviZ test suite -- we do stuff like that IIRC, and we also have built-in datasets

[AlexAndorra]

All good, thanks @ricardoV94 🍾

[AlexAndorra]

BTW @ricardoV94, do you have an email address I can contact you on please?