Save posteriors as a FITS/HDF file

[avivajpeyi]

No description provided.

[avivajpeyi]

Look into arviz

[dfm]

I think this is what we want: https://github.com/arviz-devs/arviz/blob/6589319d7bc596416602010e5e0998a57237ef17/arviz/data/inference_data.py#L215

[avivajpeyi]

Also found this: pymc-devs/pymc#2189

QS for @dfm, are we ok with

• saving the posteriors from the pymc3.backends.base.MultiTrace as a HDF5 file
• load the posterior HDF5 file as a pandas DF
• not having the loaded object be a pymc3.backends.base.MultiTrace

If we are ok with saving/loading the posteriors from a pd.DataFrame, then

posteriors = pm.trace_to_dataframe(trace)

makes it quite easy :)

[dfm]

I think the following would be a better procedure:

import pymc3 as pm
import arviz as az
with pm.Model() as model:
    x = pm.Normal("x")
    trace = pm.sample()
    data = az.from_pymc3(trace)
data.to_netcdf("chain.netcdf")

Then we could load it using:

import arviz as az
new_data = az.from_netcdf("chain.netcdf")

This has the benefit that it stores all the sampling metadata as well as the samples themselves. We could do that manually, but this is probably better!