Hi all,
I want to use the NUTS sampler to sample a model but use some form of check-pointing. I can successfully sample from the model and obtain the trace results, so that itself is not an issue.
I am working on a server, but the maximum allotted time on the server is not enough to run the model once, therefore I need to run the sample model multiple times and let it continue to run until I have sufficient samples.
Therefore, I want to be able to:
- Sample for example 1000 tunes and save the trace result
- Use the trace result to sample from further for 1000 draws and save a second trace result
- Use the second trace result and draw 1000 more samples, and save that trace result
etc. for as many draws as I need.
What would be the best approach for this? Is this feasible on PyMC alone? Any help would be appreciated.
As far as I could see, the most updated question on this was posted here, which still has not been resolved.