ClasswiseWrapper double prefixes with self.prefix when enclosed in a MetricCollection

[anivegesana]

🐛 Bug

#1866 double prefixes with self.prefix when enclosed in a MetricCollection. This is because MetricCollection already handles prefixing, here:

torchmetrics/src/torchmetrics/collections.py

Lines 335 to 339 in a448ad3

	for key, v in res.items():
	if hasattr(m, "prefix") and m.prefix is not None:
	key = f"{m.prefix}{key}"
	if hasattr(m, "postfix") and m.postfix is not None:
	key = f"{key}{m.postfix}"

but #1866 doesn't account for it.

To Reproduce

Enclose a ClasswiseWrapper with a prefix within a MetricCollection.

Finds a multiclass accuracy and a classwise f1 score.

from torchmetrics import *
import torch

category_names = ['Tree', 'Bush']
num_classes = len(category_names)

input_ = torch.rand((5, num_classes, 3, 3))
target = torch.ones((5, num_classes, 3, 3)).long()


val_metrics = MetricCollection(
    {
        "accuracy": Accuracy(task="multiclass", num_classes=num_classes),
        "f1": ClasswiseWrapper(
            F1Score(
                task="multiclass",
                num_classes=num_classes,
                average="none",
                dist_sync_on_step=True,
            ),
            category_names,
            prefix="f_score_",
        ),
    },
    prefix="val/",
)

val_metrics["precision"](input_, target)
val_metrics(input_, target)

Expected behavior

I should get {'val/acc': tensor(0.), 'val/f1_Tree': tensor(0.), 'val/f1_Bush': tensor(0.)}. I instead get {'val/acc': tensor(0.), 'val/f1_f1_Tree': tensor(0.), 'val/f1_f1_Bush': tensor(0.)}.

Environment

• torchmetrics 1.0.0 via pip
• Python 3.10.6 & PyTorch 1.12:
• OS: Linux

[github-actions]

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

[anivegesana]

Perhaps adding a flag for Metric subclasses to specify if they want MetricCollections to prefix them, like _key_names_fix = False can fix the issue.