Allow unnormalized class scores for Accuracy

[its-dron]

🚀 Feature

Presently, when using Accuracy metric on multi-class with scores (N,C entry in input types), the scores are required to be probabilities in [0, 1].

However, un-thresholded accuracy can be computed without normalized probabilities as inputs, as relative ordering of scores is all that is needed.

Given that some uses of Accuracy do require normalized probabilities, we could implement this as a flag that would disable the input check.

Motivation

It is common to work with unnormalized class scores during training, especially during classification tasks, as they are used in the more-stable nn.CrossEntropyLoss. Rather than having to additionally compute a softmax just for the accuracy metric, it would be reasonable to allow usage of arbitrarily scaled input data.

I specify Accuracy because it is the use case that I ran into, but it's possible other Metrics have the same property.

Pitch

Add a flag to Accuracy (and any other applicable metrics) that disables the input range check for preds.

Alternatives

The present workaround is to apply a softmax before feeding data to your Accuracy metric.

Additional context

https://github.com/PyTorchLightning/pytorch-lightning/blob/0456b4598f5f7eaebf626bca45d563562a15887b/pytorch_lightning/metrics/functional/accuracy.py#L25

[Jumperkables]

This please

[github-actions]

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

[jspaezp]

Just my opinion but I feel like if this is implemented, it should come with a warning the first time it happens. Since sometimes you (I definitely) would like the metric to be calculated with the current behavior.