FSC Re-scaling Issue


I have an issue with my 3D refinements. Before the implementation of V3.3.1, all of my FSCs would be calculated at full box resolution in this current or other datasets. However, this has stopped for both NU and homogeneous refinements as of the update. It now re-scales them to half of the extraction resolution. Log reads as follows:

[CPU: 1.81 GB] ====== Starting Refinement Iterations ======

[CPU: 1.81 GB] ----------------------------- Start Iteration 0

[CPU: 1.81 GB] Using Max Alignment Radius 11.520 (30.000A)

[CPU: 1.81 GB] Auto batchsize: 1946 in each split

[CPU: 2.31 GB] – THR 0 BATCH 500 NUM 500 TOTAL 8.5302715 ELAPSED 40.457999 –

[CPU: 3.51 GB] Processed 3892.000 images in 41.785s.

[CPU: 3.77 GB] Computing FSCs…

[CPU: 3.77 GB] Using full box size 320, downsampled box size 160, with low memory mode disabled.

[CPU: 3.77 GB] Computing FFTs on GPU.

[CPU: 4.02 GB] Done in 1.470s

[CPU: 4.02 GB] Using Filter Radius 35.274 (9.798A) | Previous: 11.520 (30.000A)

[CPU: 5.00 GB] Non-uniform regularization with compute option: GPU

[CPU: 5.00 GB] Running local cross validation for A …

[CPU: 7.65 GB] Local cross validation A done in 492.024s

I think this is intended? As your resolution is nowhere near Nyquist, I assume this is done to speed up FSC calculation

Ok, so then I would expect this downsampling implementation to longer be implemented once a critical GSFSC value is attained for an iteration?

That would be my expectation but maybe @team can clarify what is going on here? from the message it kind of looks like it is doing both, given it says “using full box size” and “using downscaled box size”?

Hi @singularity91 @olibclarke,

The FSC is always computed over wavelengths ranging the whole Fourier spectrum of the volume, and volumes aren’t downsampled. The log printout is meant to be used for debugging, but the meaning is a bit ambiguous.
To clarify what it means: the “downsampled” size here relates to how the masks are applied to the volume. Each FSC curve plot in a normal job shows various FSC curves with different masks applied. It’s expensive to store each mask to the full box size, so masks are computed at the “downsampled box size” and then multiplied by the (full size) volume with interpolation as needed.



Thanks for the swift responses guys. I figured something like that might be happening but I wanted to be certain that it wasn’t a sign of something more serious because was not present before. Kind regards, Lynn.