Is this strategy used for both training/testing or just for testing?

[shahzad-ali]

Thanks for sharing this repo.

1. Is this "overlap-tile strategy" is used only for inference (test time)? I'm confused even after reading the original paper.

2. Is the loss calculated for the valid part of the patch (after removing the mirror padding)?

I found useful explanation [Overlap-tile strategy in U-Nets] but it also doesn't say much on this.