sample_posterior_predictive from InferenceData?

[michaelosthege]

Description of your problem

In some scenarios, sampling and prediction do not happen at the same time or even in the same process.
sample_posterior_predictive can be fed with a MultiTrace, but in many scenarios, the trace is saved via arviz.to_netcdf and only the arviz.InferenceData is kept.

With this small function, one can convert an xarray.Dataset (the InferenceData.posterior) to a list of points that may be consumed by sample_posterior_predictive.

def _to_points(posterior:xarray.Dataset) -> typing.List[dict]:
    samples = {
        vn : posterior[vn].values
        for vn in posterior.keys()
    }
    points = []
    for c in posterior.chain:
        for d in posterior.draw:
            points.append({
                vn : s[c, d]
                for vn, s in samples.items()
            })
    return points

The list of points is probably not the most performant option. Maybe sample_posterior_predictive could be modified to consume the InferenceData object directly?

[rpgoldman]

The list of points is probably not the most performant option. Maybe sample_posterior_predictive could be modified to consume the InferenceData object directly?

Yes, for the vectorized version, I am already redoing the input away from being a set of points.

We could definitely consume the InferenceData object more directly, but sample_posterior_predictive needs access to the pm.Model object, so it's not sufficient to have a save and restore solution that is limited to the posterior trace alone.