fjug
commented
4 years ago Lovely! Additionally we should also notify people doing multi-channel data. Plus: it would be nice to test if the undertrained halo-issue become less pronounced on isotropic data. (A good picture in my mind is that a z-stack with massive \delta z is essentially the same as multiple channels between which changes are typically also not smoothly changing. Does this make sense? (No need to comment if it does...)
The information about the number of epochs used for proper training should be highlighted (red and large font).
In the 3D example we should additionally explain that more training-epochs could significantly improve the result.