Prior predictive keyword arg not used in from_mcmcchains

[bgroenks96]

When using from_mcmcchains(chain; prior_predictive=prior_pred_chain), the prior predictive chain is not stored in the resulting InferenceData.

Looking at mcmcchains.jl, it's pretty clear why; the keyword arg is listed in the documentation but I don't see it handled anywhere in the code?

[sethaxen]

Thanks for the issue! from_mcmcchains does handle the prior_predictive group, but it only does so if prior is also provided, which is not right:

ArviZ.jl/src/mcmcchains.jl

Lines 254 to 271 in 17f2e7e

	if prior !== nothing
	pre_prior_idata = convert_to_inference_data(
	prior;
	posterior_predictive=prior_predictive,
	library=library,
	eltypes=eltypes,
	kwargs...,
	)
	prior_idata = rekey(
	pre_prior_idata,
	Dict(
	:posterior => :prior,
	:posterior_predictive => :prior_predictive,
	:sample_stats => :sample_stats_prior,
	),
	)
	concat!(all_idata, prior_idata)
	end

I'll push a fix, though it might wait until after the holidays. For now, you can make it work by providing also the prior chains, which I assume you also have.

[sethaxen]

@bgroenks96 the latest ArviZ release should resolve this issue. Please let me know if it persists.

[bgroenks96]

Ok will do. I temporarily stopped using the package, though, because I ran into difficulties with the interface to InferenceData. We really need a better way to handle xarray from Julia.

[sethaxen]

We really need a better way to handle xarray from Julia.

I agree. Unfortunately, wrapping xarray really well essentially involves writing an XArray.jl package that wraps all xarray types, which is beyond the scope of this package, especially since we're working to drop Python as a dependency (see e.g. #128, #130). Unfortunately, there's really no clear best xarray drop-in alternative in the Julia ecosystem yet.

If you have the time, issues documenting the specific problems would be very useful, because I might be able to add simple workarounds that would make this package more useful in the interim.

[bgroenks96]

I would also be happy with a pure-Julia InferenceData scheme. But I doubt you have time to re-implement the multitude of functionality in xarray, though! I certainly don't...

I actually wanted to take a crack at creating an XArray.jl packge. I already have some code lying around which half-implements a wrapping scheme. I hope that I will soon have some time to work on this... I am really not happy with and am constantly bothered by the limitations and instability of DimensionalData.jl, the closest Julia alternative to xarray, at the moment.