Feature request: option for more frequent/regular checkpoints

[ajdittmann]

For context, one of the main clusters I have access to places an unfortunate limit of 12 hours on each job submission, and another has a 16-hour limit. Because my log likelihood (already written in C) evaluations take a few seconds, as I have increased N_active and N_effective into the 10,000s in search of converged results (even in this limit, pocomc seems quite a bit more efficient than many other samplers), things have gotten to the point where the jobs can time out before making it to the next iteration and dumping a pickled output file (using save_every = 1).

Would it be possible to add an option along the lines of saving the sampler state after N likelihood evaluations, or some other more predictable metric?

[minaskar]

Unfortunately, this would require major changes in the code.

Could you perhaps decrease n_active, while keeping n_effective as high as your problem requires?

[ajdittmann]

Sorry to hear that, although it is understandable. Thank you for the suggestion, I will give it a shot.