MInimum number of online-EM iterations in 2D classification

must the number of online-EM iterations be calculated so that all the particles of a large dataset (still full of garbage particles) are fully classified during these, or the final full pass would take care of those particles remaining, previously not classified, while still polishing?
For instance, to classify one million particles into 200 classes, with batchsize = 100, should I do a minimum of 50 of online-EM iterations + one final full pass, or fewer iterations + one final full pass would be enough?
Could you please advise on this?
Thank you in advance,

Fewer iterations (or default) would be fine. All particles are always seen in the last iterations. You can also add more “full” iterations at the end if you want, but these generally help only for very small or very poor SNR data.

1 Like