Resume a `cmdstanpy.CmdStanModel.sample()` run into a `cmdstanpy.CmdStanMCMC` object after finishing the python execution

[jmaronas]

Summary:

Is there a way to resume a cmdstanpy.CmdStanModel.sample() run into a cmdstanpy.CmdStanMCMC object after finishing the python execution?

Description:

After running a cmdstanpy.CmdStanModel.sample() in a e.g cluster, one would like to be able to obtain the associated cmdstanpy.CmdStanMCMC from the files generated during sampling, once these files are copied from the cluster into the local machine.

In this way, one can use e.g arviz to obtain the same inference data from the CmdStanMCMC object that we would obtain if an arviz.inference object is created just after the .sample() execution finishes. At the moment, one can only load files generated through cmdstanpy by using arviz to load cmdstan files (and in this way we miss some of the attributes that are load when directly loading a CmdStanMCMC into arviz).

Also, another (more interesting) user case would be if we would like to resume the sampling, from the last element in the chain and using the already precomputed parameters during warm up (which is usually the most time consuming step).

Current Version:

0.9.68

[mitzimorris]

a few workflow questions: what is the use case? continuing warmup? or getting more draws in order to increase the precision of your estimates - in which case, we should offer guidance on just how much precision is possible and what needs to be done w/r/t csv output config in order to get this.

In order to start a bunch of chains to continue sampling, you need the following inputs:

• the data input file - right now, input data is a .json file in the TMP dir that gets deleted at end of session
• per-chain stepsize and mass matrix
• initializations for all parameters from all chains with the parameter value for the last draw
• if the seed was specified, you'd want to use that seed and advance the RNG out far enough - cf comment here: Parallel Run Adaptive Sampler stan#3028 (comment); certainly you'd want to do this for continuing warmup. for getting more draws, a different seed would be OK too; just not the same seed and offset (chain id) as was used previously.

if you want to get more draws for more sig figs in your estimate, then we need a way to create a combined runset over all the samples. if all samples have the same number of draws, this should fly - we might want to relax the checks on the CSV header which might be overly strict w/r/t config.

[jmaronas]

The use case would be just being able to restart whatever state your model is at. So for example I could be running my model in a computer cluster for days, and suddenly if administrators say they have to shutdown the cluster for whatever reason, then I would like to be able to resume and continue in another machine.

Also, as I said in my first post, when loading cmdstan files into arviz, one gets fewer elements than when loading a CmdStanMCMC object. Perhaps those extra arguments are redundant but I think it is nicer to be able to resume exactly the same thing after sampling, than if one has the files generated by cmdstanpy stored in the hard drive.

It could also happen that the suggestion I made through this issue is a stupid suggestion :).

[mitzimorris]

The use case would be just being able to restart whatever state your model is at. So for example I could be running my model in a computer cluster for days, and suddenly if administrators say they have to shutdown the cluster for whatever reason, then I would like to be able to resume and continue in another machine.

thanks for the background.

I don't think your request is stupid - just that it covers a lot of different situations. for the particular use case - clusters - that's a bit tricky, what with node and FS setup.

[WardBrian]

If I understand the request correctly, this seems like it would be a core cmdstan feature more than an interface request, correct?

[mitzimorris]

yes and no.

core cmdstan already provides the ability to start with a specified step size, metric, and initial parameter values.
while the PRNG seed is known, we don't know how many times the PRNG was called since initialization.
furthermore, the process of parameter initialization may lead to slightly different initial parameter values than were specified.

this means that restarting won't produce identical results to runs which ran for the same number of iterations without the 't stop/restart.

if users are OK with this, then the interface should be able to do the output file munging / input data assembly. it would require keeping around the input data files as well as the CSV output files.

[WardBrian]

This was done in the R ecosystem with its own package, https://donaldrwilliams.github.io/chkptstanr/. I think we may similarly want to say this is out of scope for the basic cmdstanpy wrapper

[mitzimorris]

agree - not going to do this in CmdStanPy.