Accuracy Metric does not perform squeeze on input data

[jungla88]

Hello,

I am not sure if this is a bug or not but according to https://torchmetrics.readthedocs.io/en/stable/pages/classification.html#input-types, prediction tensor is squeezed starting from dimension one. Accuracy seems not performing such operation and in validity checks raised an error:

preds = torch.tensor([0.4,0.51,0.51,0.6,0.7])
target = torch.tensor([0,1,0,0,1])
acc = Accuracy(num_classes=1)
acc(preds,target)

This works ok, whilst when acc(preds.unsqueeze(dim=1),target):

ValueError: The highest label in target should be smaller than the size of the C dimension of preds.

[Borda]

@jungla88 — this was a real issue with the old unified Accuracy class but has been resolved by the task-specific API refactor. As of v1.9.0:

Use BinaryAccuracy for binary classification:

from torchmetrics.classification import BinaryAccuracy

acc = BinaryAccuracy()

preds = torch.tensor([0.4, 0.51, 0.51, 0.6, 0.7])
target = torch.tensor([0, 1, 0, 0, 1])

# Works with 1D
acc(preds, target)  # tensor(0.6000)

# Also works with extra dim — internally handled
acc(preds.unsqueeze(1), target)  # same result, no error

For spatial predictions (segmentation):

acc = BinaryAccuracy(multidim_average="global")   # flatten all spatial dims
acc = BinaryAccuracy(multidim_average="samplewise")  # per-sample accuracy

The old Accuracy(num_classes=1) pattern is deprecated. Always use task-specific classes:

Task	Class
Binary	BinaryAccuracy
Multiclass	MulticlassAccuracy(num_classes=N)
Multilabel	MultilabelAccuracy(num_labels=N)

Docs: Accuracy