Unqualified threshold checks contradict logits support

[MichaelSpece]

🐛 Bug

Accuracy both functional and module-wise are currently designed to accommodate logits. However, the most natural choice of a threshold (0) fails in both cases. More generally, thresholds outside of (0, 1) trigger ValueError. This is documented behavior that contradicts logits support.

To Reproduce

Code sample

torchmetrics.functional.classification.accuracy(torch.tensor([0]), torch.tensor([1]), threshold=0)

or

torchmetrics.classification.accuracy.Accuracy(threshold=0)

Expected behavior

Ability to choose arbitrary float thresholds without triggering an exception.

Environment

Bugs are apparent from the source code itself. The threshold checks are unqualified.

Additional context

Predictions are formed from a given preds via preds >= threshold.

Threshold checks were recently removed for F1 (#350, #351) to support logits.

Open to assignment.

[github-actions]

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

[SkafteNicki]

Fixed in PR #351, please update to master:

pip install git+https://github.com/PytorchLightning/metrics.git@master

Closing issue.

[MichaelSpece]

Hi @SkafteNicki , Apologies for the confusion, but could we reopen this issue? I was working off the latest (post-#351) master, as alluded to in my context description. Please see remaining offending code at the following:
https://github.com/PyTorchLightning/metrics/blob/a5af6d851b4236483a3f34f15ccfa79bb90a728a/torchmetrics/classification/accuracy.py#L142-L143
https://github.com/PyTorchLightning/metrics/blob/a5af6d851b4236483a3f34f15ccfa79bb90a728a/torchmetrics/classification/accuracy.py#L208-L209
https://github.com/PyTorchLightning/metrics/blob/a5af6d851b4236483a3f34f15ccfa79bb90a728a/torchmetrics/functional/classification/accuracy.py#L244-L245
https://github.com/PyTorchLightning/metrics/blob/a5af6d851b4236483a3f34f15ccfa79bb90a728a/torchmetrics/functional/classification/accuracy.py#L274-L275
Logits support (with thresholds) is claimed here:
https://github.com/PyTorchLightning/metrics/blob/a5af6d851b4236483a3f34f15ccfa79bb90a728a/torchmetrics/classification/accuracy.py#L56
https://github.com/PyTorchLightning/metrics/blob/a5af6d851b4236483a3f34f15ccfa79bb90a728a/torchmetrics/functional/classification/accuracy.py#L210

[SkafteNicki]

Hi @MichaelSpece, thanks for making me aware that the code indeed is not updated everywhere.