Accuracy metric - negative ignore_index fails when combined with subset_accuracy=True

[antifriz]

🐛 Bug

To Reproduce

Combine ignore_index=-1 with subset_accuracy=True

Code sample

import torch
import torchmetrics

accuracy = torchmetrics.Accuracy(
    subset_accuracy=True,
    ignore_index=-1,
)
# element with index 2 will always be outputted
preds = torch.arange(45, dtype=torch.float).reshape(3, 5, 3).permute(0, 2, 1)
target = torch.tensor(
    [
        [2, 2, 2, 0, -1],
        [2, 2, -1, -1, -1],
        [2, 2, 2, 2, 2],
    ],
    dtype=torch.int64,
)
accuracy.update(preds=preds, target=target)

Expected behavior

Not crashing

Environment

• OS (e.g., Linux): OSX
• Python & PyTorch Version (e.g., 1.0): 3.9 (torch==1.11.0, torchmetrics==0.8.0)
• How you installed PyTorch (conda, pip, build command if you used source): pip
• Any other relevant information:

Additional context

Stack trace

Traceback (most recent call last):
  File "main.py", line 18, in <module>
    accuracy.update(preds=preds, target=target)
  File ".../python3.9/site-packages/torchmetrics/metric.py", line 312, in wrapped_func
    update(*args, **kwargs)
  File ".../python3.9/site-packages/torchmetrics/classification/accuracy.py", line 233, in update
    correct, total = _subset_accuracy_update(
  File ".../python3.9/site-packages/torchmetrics/functional/classification/accuracy.py", line 224, in _subset_accuracy_update
    preds, target, mode = _input_format_classification(
  File ".../python3.9/site-packages/torchmetrics/utilities/checks.py", line 432, in _input_format_classification
    target = to_onehot(target, max(2, num_classes))  # type: ignore
  File ".../python3.9/site-packages/torchmetrics/utilities/data.py", line 99, in to_onehot
    return tensor_onehot.scatter_(1, index, 1.0)
RuntimeError: index -1 is out of bounds for dimension 1 with size 3

[github-actions]

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

[SkafteNicki]

Issue will be fixed by classification refactor: see this issue #1001 and this PR #1195 for all changes

Small recap: This issue describes that using an negative ignore_index fails when combined with subset_accuracy=True. After the refactor, accuracy will no longer support the subset_accuracy argument. Instead this functionality have been split into its own metric called exact_match. This metric (together with all other classification metrics) now support negative ignore_index arguments.

from torchmetrics.functional import multilabel_exact_match
import torch
preds = torch.tensor([[0.11, 0.22, 0.84], [0.73, 0.33, 0.92]])
target = torch.tensor([[0, 1, -1], [1, 0, 1]])
multilabel_exact_match(preds, target, num_labels=3, ignore_index=-1)

which works and gives the correct result.
Issue will be closed when #1195 is merged.