Does 3D prediction fail if the Z dimension is too small (not a power of 2, less than 8)?

[jdeschamps]

Question

We should test whether images with low number of slices (e.g. shape (5, 128, 128)) fail when predicting because of the Z dimensions not being power of 2 and less than 8?

If they do, would padding be a work-around?

[jdeschamps]

Also downsampling is applied similarly in Z as in YX, which might be suboptimal given that dimensions and resolutions are often different in Z.

[mese79]

Yes, It will fail if the Z dimension of a test stack is less than 8.
Especially for the BMZ export which calls the predict with the default config params (link).

I'm wondering if this is the model (algorithm) limitation, or just the CAREamics api ?