Repository navigation
easy way to limit sampling to a maximum time duration #474
Description
Activity
Thanks for the feature request. Seems like a reasonable request to me.
What do you envision should happen after the sampling is canceled? Should the fit and CSV file be discarded or should it be possible to read in the samples?
Should the fit and CSV file be discarded or should it be possible to read in the samples?
I'm am against discarding anything automatically.
What do you envision should happen after the sampling is canceled?
For my application, I just need to know what happened so I can record the error.
Reacted by Rok ČešnovarRelated: #426
Reacted by Rok Češnovar and Joshua PritikinUntil cmdstanr gets some improvement, here's a solution:
library(R.utils) fit <- try(withTimeout(m1$sample(dl, chains = 4, parallel_chains = 4, iter_warmup = totalIter/2, iter_sampling = totalIter/2, # metric="diag_e", inv_metric = im.prior, init=function(chain_id) list(B=0,eVar=.95), output_dir = localDir), timeout=60*60, cpu=Inf), silent = TRUE) if (is(fit, "try-error")) { recFail(paste("Timeout during sampling")) }Reacted by Rok Češnovar, Mike Lawrence and Jonah Gabry
Is your feature request related to a problem? Please describe.
I have a model that often runs in a reasonable amount of time, but occasionally, when faced with unpleasant data, will continue to sample for 4+ hours. Rather than wait, I'd prefer just to record the failure and move on to the next dataset.
Describe the solution you'd like
I'd like to be able to bound the maximum time duration for sampling.
Describe alternatives you've considered
Maybe there is some R package that wrap cmdstanr's sampling API and offer this functionality? Oh, hm, I found withTimeout in the
R.utilspackage.