Coordinates and values of `constant_data` messed up

[michaelosthege]

Description of your problem

I have a model that has two data variables of identical dimensionality:

pymc3.Data("obs_A360", obs_A360, dims=("replicate_id", "cycle"))
pymc3.Data("obs_A600", obs_A600, dims=("replicate_id", "cycle"))

where obs_A360 and obs_A600 are numpy arrays of shape (131, 5).

Sampling works fine, also an InferenceData object is returned.

But looking into idata.constant_data the shapes, values and coordinates of those data variables are completely messed up:

• The replicate_id coordinate became a concatenation of the string-valued replicate IDs with a numpy arange of the same length.
• idata.posterior.replicate_id and idata.constant_data.replicate_id don't match
• constant_data.obs_A360 and constant_data.obs_A600 have shapes (262, 5) and are half filled with nan. And not even the same way!

What's going on?!

Versions and main components

• PyMC3 Version: main

[OriolAbril]

This might be an issue with the faulty workaround from #5042. Because it is triggered only when default_dims is not None -> not triggered for posterior which does work, but triggered by constant data and all hell gets loose.

I'd recommend fixing #5042 and then check to see if the issue persists.