Fix small bug

[ColCarroll]

This looks like a typo, and probably caused funny slicing problems when sampling twice from a model in the past. draws_idx is not used anywhere else in the project.

[junpenglao]

Good catch!
I guess this is for when you continue sampling and stacking the sample to an old trace? didnt realized this is possible.

[ColCarroll]

Yep. I believe you can do something like

with pm.Model() as model:
    ...
    trace = pm.sample()

...
with model:
    trace = pm.sample(trace=trace)

I spotted it while working on a tiny utility to allow

with pm.Model() as model:
    ...
    trace = pm.sample()

pm.save_trace(trace, 'eight_schools.trace')
...
with model:
    trace = pm.load_trace('eight_schools.trace')

Initializing the trace needs to have access to the model for historical reasons...

[junpenglao]

Cool!
Is this recommended? My biggest question to it is that since the previous trace is after tunning, and I suppose the next trace will start from 0 and tune again there is a risk that the chain is not in equilibrium any more? Maybe a warning should be print for this usage...

[ColCarroll]

Yes - I do not think this use case has been thought about/maintained in a long time - serializing tuning data seems difficult.

Even for saving/loading the trace, the use case I am considering is a Jupyter notebook, where I would like to run the analysis once, but be able to reload the posterior samples later, perhaps on another machine.