cube_size vs. crop_size during refinement

[Simon Wiedemann]

Dear IsoNet Team,

First of all, I highly appreciate your software package and how easy it is to use it. 

I recently started experimenting with IsoNet, and there is one thing that I cannot quite wrap my head around. It is the difference between cube_size and drop_size during the refinement step. From the documentation, I understand that crop_size is the size of the extracted subtomograms, while cube_size is the size of the data used to actually train the model. The documentation says that crop_size should be larger than cube_size. It is clear that crop_size > cube_size is important for the prediction step, as it means more overlap, reducing tiling artifacts. 

My questions are:

1. Why is crop_size > cube_size is important for model training?

2. Assume I used another software package to split my tomogram into chunks (crops) instead of IsoNet's extract command. Assume the subtomograms have the shape NxNxN. Is it then safe to set cube_size = crop_size = N when training a model on these subtomograms?

I would highly appreciate any thoughts on these two questions.

Best,

Simon

[YUNTAO LIU]

Hi Simon,

You are correct, crop_size>cube_size is not important in refine step.

There are two concerns when we designed this larger crop_size. First to avoid the edge effect when doing Fourier transform. Second we planed to implement random rotations, the larger crop size will reduce blank areas after rotation. 

--
You received this message because you are subscribed to the Google Groups "IsoNet" group.
To unsubscribe from this group and stop receiving emails from it, send an email to isonet+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/isonet/43ae6258-cad8-440a-8272-79f1bab8cc10n%40googlegroups.com.

--

Best Regards,

Yuntao Liu,  Postdoc.

California NanoSystem Institute

University of California Los Angeles