It is very evident from inspection of patch & rigid motion trajectories that beam induced motion is heterogeneous within a dataset - some exposures have fast motion at the beginning, while others do not, likely relating to ice thickness. But during Patch Motion and RBMC, one set of dose weights is used for all the micrographs.
To account for this, I wonder if it would be possible to extend the empirical dose weighting scheme in RBMC to fit per-micrograph empirical dose weights, in cases where the particle density and resolution justify it?
This is not a totally new idea, it has been implemented in modifications of FREALIGN with apparently good results (paper here).
Hi @olibclarke. Interesting idea. I’m skeptical that there’s enough signal in typical datasets to accurately fit a dose weight curve per micrograph. You could try using exposure curation to split your movies into subsets based on e.g. ice thickness, and do separate dose weight computations on each subset. I’d be very interested to hear if you discover a benefit to doing this. That said, I’ll have a look at the paper you’ve linked and take this back to the rest of the team for discussion. Thanks for bringing it up.
Good idea - trying this right now, splitting a dataset into thirds based on relative ice thickness and running RBMC on each third separately.
It occurs to me that the motion correction hyperparameters might also vary between thin and thick ice
Is it possible to just compute empirical dose weights for a subset given existing per particle trajectories?
My initial though was that the motion-correction hyperparameters would typically vary more between thin and thick ice than empirical dose weights.
But then, the thick-ice movies might never contribute much to the high-resolution resolution realm anyhow in a thin-thick ice mixed dataset while they may add useful information at lower res (such as wider orientational distribution). So including them would be better than totally excluding them through an early ice filter.
Looking forward to what comes out of your experiment!
I’m trying this now - I’ve split the dataset into thirds based on ice thickness, and am performing RBMC on each third separately.
Question - is there any way to quantitatively compare the empirical dose weights generated by each job?
So far there does seem to be a difference in the motion hyperparameters, although too soon to tell whether it is significant - the spatial prior and spatial correlation distance are very similar between the thin and thick ice sets, but the acceleration prior is quite different - it is 0.14 for the thin ice set, 0.0049 for the thick ice set, as compared to 0.012 for the run where all mics were combined.
is there any way to quantitatively compare the empirical dose weights generated by each job?
If you mean to compare their impact on resolution, you can use apply trajectories followed by a refinement. If you mean compare them numerically, you can do so with the help of python. In the refmotion job folder you’ll see a file called
refm_empirical_dw.npy which you can read with
numpy.load(path) - it’ll return the actual dose weight array.
Thanks Harris! At the moment I am having the problem that the subset runs are stalling out, which is rather frustrating, so I can’t compare them - any advice appreciated: Stalled RBMC jobs? - #4 by olibclarke
Responded in the public thread - is this still an issue?
Seems to be okay after a system restart, although can’t be sure until they actually complete (will take a day or two) - thanks!