top-k multiclass macro accuracy is not calculated correctly

[markussteindl]

🐛 Bug

If some classes are not observed, the top-k multiclass macro accuracy is not calculated correctly.

To Reproduce

from torchmetrics import Accuracy
import torch

acc1 = Accuracy(task="multiclass", num_classes=3, average="macro")
acc2 = Accuracy(task="multiclass", num_classes=3, average="macro", top_k=2)

preds = torch.tensor([[0.9, 0.1, 0.0], [0.9, 0.1, 0.0], [0.9, 0.1, 0.0]])
targets = torch.tensor([0, 0, 2])

print(acc1(preds, targets))  # tensor(0.5000) --> ok
print(acc2(preds, targets))  # tensor(0.3333)  --> inconsistent

Expected behavior

In the example, the top-1 and top-2 accuracy should be equal.

Note that class 1 does not occur in the targets.
The classes have the following top-1 and top-2 accuracies before averaging:

• class 0: 1.0
• class 1: nan
• class 2: 0.0

Thus, the macro aggregation of the top-1 and top-2 accuracy should be equal with a value of 0.5 (or 0.333, if we take the nan into account).

Environment

• TorchMetrics version: 1.3.0.post0, 1.3.1
• Python & PyTorch Version (e.g., 1.0): Python 3.11, torch 2.2.0
• Windows 10, 22H2, 19045.4046
• Environment managed by poetry