Faulty logic to retrieve tune from mixed samplers

[ricardoV94]

This line here assumes that the first sample has tune information, which is not necessarily the case (as the comment above states):

pymc/pymc/sampling/mcmc.py

Lines 738 to 742 in 457421b

	# count the number of tune/draw iterations that happened
	# ideally via the "tune" statistic, but not all samplers record it!
	if "tune" in mtrace.stat_names:
	# Get the tune stat directly from chain 0, sampler 0
	stat = mtrace._straces[0].get_sampler_stats("tune", sampler_idx=0)

Example that fails (but could be fixed in a future PR without solving the underlying issue:

import pymc as pm

with pm.Model() as m:
    x = pm.Normal("x")
    y = pm.Normal("y", x)
    idata = pm.sample(step=pm.Slice([x]))  # KeyError: 'tune'

As the Slice sampler does not provide tune stats

[manulpatel]

Hi @ricardoV94! So ideally how should we get tune in general from any sampler?

[ricardoV94]

It should be possible to check which sampler contains the tune statistic. If it's too hard a loop with try/except would also work.

[ricardoV94]

Quoting @michaelosthege in #6835 (comment)

For the record: Our current approach of passing information about tuning/not-tuning via sample stats is brittle. It's a per-sampler thing, but in practice it's actually a per-iteration thing from the PyMC, ArviZ and McBackend perspective. IMO we should consider taking it out of the step method attributes and stats, and instead pass tune as a parameter to the astep method.