Hello all,
I am experiencing a strange failure mode when processing helical data. The failure appears to be caused by the ‘Force max over pose/shifts’. Using the same selected particles, one job fails with ‘Force max over pose/shifts’ switched off while the other succeeds with it switched on. All previous jobs had it switched off (which I prefer in this case) and worked well.
Any ideas on how to fix this problem and allow me to run with the force max switched off again?
Thanks!
Charlie
Force/max off takes a lot longer to converge. How many iterations did you run it for?
If you left the number of iterations at default, it won’t converge, exactly as Oli says. With Force Max off, I had one dataset which needed 120 iterations before I was happy with what I was seeing in the 2D classes. Other things to tweak as the initial sigma anneal (although 99% of the time I leave this alone), class uncertainty (I often get nice results with 5 or 8 instead of the default 2, and a second round of 2D seems to find less junk hiding in nice classes) and of course iterations (I like 3 full final iterations, seems to work nicely) and particles per class (depending on number of classes and size of dataset).
The job from the first image I killed after a few early iterations as I had never seen a job that starts this way end up in any decent classes. The settings for that killed job are:
Classes: 100
Uncertainty: 5
Force max: False
Online EM: 40
Final Full: 5
I will restart and allow this to complete now.
At an earlier stage with this pipeline, while cleaning, I had run with the below custom settings (see imaged, based on some other comments from Oli about clean up) and never saw this early iteration state. I should also say that initial iterations on this jobs were clearly just misaligned filaments rather than blobs.
Class: 100
Uncertainty: 1
Force max: False
Online EM: 40
Final Full: 5
Iteration to start annealing sigma: 200
Re-center 2D classes: False
UPDATE:
So big egg on my face here - allowing the jobs with that strange initial iterations appearance to complete yields ‘normal’ results.
Lesson learned! Thank you both for your inputs!
