Potential inconsistency between paper and implemented code

[KBlansit]

Hi. I was looking at your implimentation of CORN loss for ordinal regression. When looking at your paper on archive (the June 1 2023 version), I noticed in the supplemental materials S1 the corn loss pytorch definition does appear to match what is implemented in this repository.

More specifically, while the archive paper appears to normalize the loss as return losses/num_classes, the repository normalizes the loss as return losses/num_examples.

I was wondering the source of the discrepancy between the two. Thank you for your time and your contributions.

KB