Ordinal classification metrics

[Zhylkaaa]

🚀 Feature

Hi I would like to contribute some basic ordinal classification metrics like "accuracy within"

Motivation

I recently discovered that my problem is an instance of ordinal classification problem. Which means that classes with labels that are close, are also semantically close (like star rating 1 and 2 start are not really far apart). And I realized that torchmatrics doesn't implement for example "accuracy within k" (or off-by k), which is stricter version of top-k accuracy. More on that can be found here: https://link.springer.com/chapter/10.1007/978-3-642-01818-3_25

Pitch

I would like to contribute at least Accuracy/Recall/Precision in "within k" version, but I am not sure if I should disrupt original classification metrics by adding more parameter. It seems like logical solution, but I am not sure how much you would be open to disrupting core metrics implementations.

Essentially, e.g. true positive definition would change from (pred == label).sum() to (torch.abs(pred - label) <= within)sum()

Alternatives

I can also implement them separately or just do it for my project :)

Additional context

In my context I am doing IVF image stage classification and stages are visually very similar, because some of the stages differ only by single cell, which might be actually hidden, since its a 2d top projection of a 3d object.

[github-actions]

Hi! thanks for your contribution!, great first issue!