mathiasmorch
Hi,
I was wondering if there is a way to show/get/output the correlation matrix when using bootstrap_errors?
Sorry if this is an ill-defined question, I am still reading up on bootstrapping methods.
I'm working on some parametric/surface refinements with ~ 200 - 400 patterns and "do_errors" makes the refinement freeze, but I still think it could be interesting to inspect parameters for any egregious correlation.
Alternatively - is there a way to inspect the correlation for a pair of parameters when doing bootstrap_errors?
Also, does "local_with_error" exist in some form? I hoped doing something similar to what you can do with "prm_with_error", and keeping errors from temperature measurements.
rowlesmr
Not answering your question, but how long does it freeze for?
I've accidently had do_errors defined when doing some fairly hairy stuff, and while it looks like it's frozen, it is working away in the background. Just as a test, maybe setup a refinement to finish on a Friday afternoon, and give it the weekend to do the errors?
.
Also, your suggestion of local_with_error is a really good idea. I'm wondering if you could do a work around by defining a bunch of prm_with_errors wth unique names, and then have local nameOfChoice = prmWithErrorName; I don't know if it would work, but it would be worth investigating.
Matthew
alancoelho
Can I ask how many parameters are being refined.
Normal error calculation for a correlation matrix is good for about N=3000 independent parameters. It involves matrix inversion which is an N^3 process. Thus if you have 3000 parameters then that's an inner loop of 27 billion.
mathiasmorch
Thanks for your swift replies Matthew and Alan.
@Matthew
Maybe i should try running it over night / over weekend again, I just remember running it for a long time and giving up.
I've tried doing the prm_with_errors and then loading it into a local, but to me it doesn't seem like it carries over the error, but I might have been too hasty in that conclusion and should revisit it.
@Alan
Of course; I'm refining between 4000-7000 parameters at the moment.
It's a few different sets of data that each have between 250 and 400 patterns.
I would love to bring the # of parameters down and also have models for the same datasets with fewer parameters.
While evaluating Rwp, difference maps, errors on parameters and so on is nice, I would really like to compare correlation.
For example: anisotropic crystallite size and preferred orientation with and without parametrization of these parameters, and also with different parametrization models.
Additionally I find it likely that if I extend some studies to in-situ synchrotron experiments the # of patterns might grow even further, and i would love to be able to continue doing parametric refinements on that.