`random_size` defaulting to `False` in RandSpatialCropd

[wyli]

Discussed in #5651

^{Originally posted by MathijsdeBoer December 5, 2022}
Hello all!

I've just spent way too much time on debugging a weird shape issue (as is common in deep learning) when trying to get a quick prototype going. Eventually I found out that the input patches would have inconsistent shapes, leading me to run through all of my transforms expecting a mistake, or some other unexpected behavior that I didn't account for.

Eventually, when running through I noticed that transforms.RandSpatialCropd (And the non-dictionary version) have random_size=True by default. To my mind, this is counter to expected default behavior. If I call a RandSpatialCropd(keys=some_keys, roi_size=some_size), I am not expecting my outputs to have random shapes unless specifically enabled.

Would a switch to random_size=False be a reasonable change to make in the next update?

[surajpaib]

@wyli To add to this discussion, I also agree that this would be a reasonable default considering my use cases. This applies to transforms like RandScaleCrop as well. I've explicitly set the random_size=False for quite a few projects.

I'm happy to contribute if this is still on the roadmap to be included.