I agreed to do a review of the stantargets package for ropensci’s process. It’s been really cool to dig through a package from the reviewer perspective and I’m posting to ask for your experiences with the usability of the package. Feel free to comment/fangirl/complain here and I’ll do my best to check out the aspects of
stantargets you mention so I can include them in my review.
Porting some of my data/modeling pipelines to
stantargets has been fun but I know my own perspective on the usability aspects is going to be limited so I appreciate any pointers!
Started using it over the weekend for SBC. I got up to speed pretty quickly but am increasingly now considering abandoning it for a plain ol’ targets pipeline for more control (esp. doing proper/full SBC).
Can you say what the painpoints are?
Well, it currently relies entirely on what’s available in posterior for what it provides in tar_stan_mcmc_rep_summary, which means no info on whether a rep encountered any divergences and no ranks for SBC, which is pretty limiting. Also, when running things despite those output limitations, I ended up having to abandon the batch/rep framework and instead use tar_map to get the order of execution I wanted.
That’s a lot of helpful detail, thanks!
tagging @paul.buerkner if he has missed this comment. It would be useful if you have time to create issues also for the posterior (the dynamic HMC specific diagnostics for posterior package have already been discussed)
I overlooked this indeed, thanks.
@mike-lawrence, since you have been playing around with it, would you mind opening issues at GitHub - stan-dev/posterior: The posterior R package for what you think is missing in posterior (and hence in stantargets) currently?
It’s already there as #111
ah OK thanks. wasn’t sure if it was only that or required more stuff to be added to posterior.