Model scoring in pymc

1

Hello - I have a question on how to score a pymc model on a different dataset. Here is how I have structured my pymc model -

def model_factory(data1):
    with Model() as example_model:
    ....
    ....
    return example_model

And then to generate trace I use the following code -

with model_factory(data1):
    trace = pm.sampling_jax.sample_numpyro_nuts(
            draws=1000,
            tune=1000,
            chains=4,
            random_seed=1111,
            target_accept=0.95
    )

To generate posterior predictive samples, I use the following code -

with model_factory(data1):
    predictive_preds = pm.sample_posterior_predictive(trace=trace, random_seed=1234).to_dataframe(include_coords = False)

My question is - How to generate samples on a different dataset(data2) while using the trace which has been generated using data1? I have been just feeding the “with model_factory” stack with a data2 like this -

with model_factory(data2):
    predictive_all_zeroed = pm.sample_posterior_predictive(trace=trace, random_seed=1234).to_dataframe(include_coords = False)

Is this the right way of scoring the model on a different dataset which was not used for trace generation?

2

Have you seen the Out-Of-Sample Predictions notebook? If not, I would probably suggest starting there.