Fix sample-counting and resuming bugs #41
Conversation
from the previous chain thin times. Give an error unless the old chain has a number of samples that is a multiple of isave/thin.
Write out the initial sample to the chain file and then do Niter iterations of the MCMC process. If Niter = a*isave + b*thin + c, the number of samples in the output will be a*isave/thin + b plus the initial sample. Only allow resuming when the file contains only complete rows and their count is one plus a multiple of isave/thin. Raise an exception if isave is not a multiple of thin. Progress reports include percentage of post-resume work accomplished. Fix documentation of metaparameters following Paul Baker
|
I tested many cases involving various values of |
Better error message trying to resume with partial blocks.
|
There was a bug involving parallel tempering in which partial blocks did not get written out except in the cold chain. Now that is fixed and I have tested it with multiple threads and everything seems to work. |
|
@kdolum, the works looks good. However, the integration tests are failing at the moment for reasons not related to your PR. I think it's good policy to fix those first before merging. Let me first have a look at that first. You also write you haven't been able to test with MPI. Is that still correct? I saw your message regarding thermodynamic integration the other day, so perhaps you have been looking into this. The tests do run locally on my workstation, so that's at least a good sign. |
|
I did test it with MPI and all seems well. |
|
This PR depends on #44 to pass the integration tests. So that one first needs to be handled before we can merge this PR |
compatibility with updated flake8
Put a long format string in parentheses to make a line shorter and satisfy flake8.
| # bare except ok | ||
| E722 | ||
| # lambda expressions ok | ||
| E731 |
There was a problem hiding this comment.
I didn't need to do these edits. But if they gave you problems locally, I don't see a reason to decline this
Fix bugs #10 and #20 as described in #10. In the usual case where
Niteris a multiple ofisave, all computed samples will be written out. The run can be resumed with some largerNiterand the samples will be correctly counted. IfNiteris a multiple ofthinbut not ofisave, the partial block of samples will be written at the end. In this case you cannot resume without deleting this partial block by hand. IfNiteris not a multiple ofthin, the extra samples will be computed but not written.ofProgress reports after resuming give the percentage of the new samples that have been computed, as well as the percentage of the total.