provide checkpointing via checkptstanr

[jsocolar]

https://cran.r-project.org/web/packages/chkptstanr/vignettes/chkpt_brms.html

[venpopov]

Seeing this issue reminded me that I wanted to the same for our bmm package. Turns out it was pretty straighforward (at least for our infrastructure) - you could check out here, if you are interested (although it requires a bugfixed version of chckptstanr that is not on CRAN).

Update: I just talked to the original developer of chkptstanr, and he said I can take over the package maintenance and development. So let me know if I can implement anything to make this easier for your package (feel free to open any issues and feature requests at https://github.com/venpopov/chkptstanr)

[venpopov]

I actually don't recommend trying it yet. After emailing with the original creator, I agreed to take over the development, and have found a lot more bugs. I'm planning to fix them, but the code needs substantial refactoring and it will take a while (though I think it's worth it - even with the remaining bugs, playing around with it for the last few dayhs and adding a few new features has been a lot of fun). I have an initial roadmap for development here: venpopov/chkptstanr#1

[jsocolar]

@venpopov I imagine you have everything under control, but in case it's useful I understand Stan's adaptation schedule reasonably well and would be happy to provide input/review of your chkptstanr work.

[venpopov]

That would be awesome. Do you want me to add you to the repo? Happy to have you either in just a review capacity, or if you want to contribute any code, whichever fits your interests/time.

Where we currently stand

The original package has two components:

• the logic of the checkpoint and the procedure of accomplishing it
• the code for implementing it

The first one seems (with one exception), generally good and working* and I get similar results with checkpointing as with normal sampling. The code, however, is a huge mess. In that regard:

• I have fixed most major bugs. Only one remains, which turned out to be the biggest mess (Incorrectly calculated number of draws in the brmssummary causes various postprocessing methods to fail venpopov/chkptstanr#8). I think I should have that ready in a day or so
• I want to refactor the code fundamentally to make the work much simpler

If you'd like to join in, the most reasonable point would be after the refactoring - then the code should be much cleaner, easier to follow, understand and test.

--
* by working I mean that I ran a lot of different models of different complexity with and without checkpointing and I generally get very similar results