Often we will start using blob picker in Live, then once we have a few decent 2D classes from the streaming job (not in template generation, just regular streaming 2D), we use these to re-pick using template picker and re-extract.
How are these newly re-extracted particles treated for the purposes of 2D? It seems like they are by default correlated with the existing 2D references, is this the case? Are the existing particles out of 2D automatically discarded when there is a change in extraction parameters?
The reason I ask is that often after doing this the streaming classes seem to degrade - but the picking is genuinely better, and indeed the classes improve markedly when I restart with a different number of classes (to force it to start from scratch).
When you switch your picker from the blob picker to the template picker, the Streaming 2D Classification job will “start from scratch”. This will throw out any existing accumulators and start fresh, similar to when you first run the job. Other ways that the Streaming 2D Classification will “start from scratch” is when the “Extraction box size” or “Fourier crop to box size” parameters in the Extraction stage change, or the “Output F-crop factor” parameter in the Motion Correction stage changes. You can force this to happen by hitting the “Restart” button, which will force the job to “start from scratch” after the next time it enters “Waiting” status.
From my understanding of your post, it seems like the Streaming 2D Classification job doesn’t actually restart when you switch pickers. Is it possible if you can show me an instance of this happening by pasting the log events from the Streaming 2D Classification job where this happened?
correct, that definitely does not happen in my hands. Yes, the next time I come across this I will post logs. I have had similar issues with changing box size - it behaves weirdly and gives bad classes until I restart the streaming 2D with a different number of classes, which does seem to force it to start over properly.
Ok, so I just tried this now, and the same thing happened again. After changing template picking parameters and re-extracting, in the log, it says it is starting from scratch, but the resulting classes look lousy. Whereas if I change the number of classes slightly and hit restart, the classes quickly get very good.
We’ll look into this. For the time being, please restart your Streaming 2D Classification job whenever you change pickers. Sorry for the inconvenience.