Inconsistent default average value in F1Score (and others) compared to MultiClassF1Score

[OmerShubi]

🐛 Bug

The default value for the average field in F1Score is micro:

torchmetrics/src/torchmetrics/classification/f_beta.py

Line 1107 in fb84f75

average: Optional[Literal["micro", "macro", "weighted", "none"]] = "micro",

However, the default in MultiClassF1Score is macro:

torchmetrics/src/torchmetrics/classification/f_beta.py

Line 785 in fb84f75

average: Optional[Literal["micro", "macro", "weighted", "none"]] = "macro",

When using the F1Score function, it seems as though that micro will be used by default, which is misleading.
In practice, when using F1Score with task='multiclass' the default value is macro.

*The same inconsistency exists in Precision and Recall (and maybe others).

To Reproduce

See F1Score / MultiClassF1Score function calls.

Expected behavior

Maybe delete the default value in the F1Score wrapper, as it is anyways overridden?

Environment

• TorchMetrics version (and how you installed TM, e.g. conda, pip, build from source): 0.11.4, conda
• Python & PyTorch Version (e.g., 1.0): 3.11.3, 2.0.1
• Any other relevant information such as OS (e.g., Linux): Linux

[github-actions]

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

[OmerShubi]

I was using an old version of torchmetrics.
In 1.1.1 the function parameters in the F1Score function (and others) are not shown in the editor (VS Code), but the misleading default value remains.

[Borda]

but the misleading default value remains.

@SkafteNicki mind check it, pls

[SkafteNicki]

@Borda unsure what to do about this. Essentially, we cannot fix it without breaking some kind of backwards compatibility.
The reason that base classes F1Score, Precision, Recall etc. have the average set to micro is that was the default value before the classification refactor. After the classification refactor and the introduction of the class specific metric MulticlassF1Score, MulticlassPrecision, MulticlassRecall we set the default to macro. So the reason for this difference is that we tried to keep backward compatibility to metrics before the refactor.