Merge branch 'master' into feature/multioutput_mse

Author: mergify[bot]

CHANGELOG.md

@@ -41,6 +41,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
+- Fix support for int input for when `multidim_average="samplewise"` in classification metrics  ([#1977](https://github.com/Lightning-AI/torchmetrics/pull/1977))
+
+
 - Fixed x/y labels when plotting confusion matrices ([#1976](https://github.com/Lightning-AI/torchmetrics/pull/1976))
 
 

src/torchmetrics/functional/classification/stat_scores.py

@@ -374,10 +374,10 @@ def _multiclass_stat_scores_update(
             preds_oh = torch.movedim(select_topk(preds, topk=top_k, dim=1), 1, -1)
         else:
             preds_oh = torch.nn.functional.one_hot(
-                preds, num_classes + 1 if ignore_index is not None and not ignore_in else num_classes
+                preds.long(), num_classes + 1 if ignore_index is not None and not ignore_in else num_classes
             )
         target_oh = torch.nn.functional.one_hot(
-            target, num_classes + 1 if ignore_index is not None and not ignore_in else num_classes
+            target.long(), num_classes + 1 if ignore_index is not None and not ignore_in else num_classes
         )
         if ignore_index is not None:
             if 0 <= ignore_index <= num_classes - 1:

tests/unittests/classification/test_stat_scores.py

@@ -542,6 +542,15 @@ def test_multilabel_stat_scores_dtype_gpu(self, inputs, dtype):
         )
 
 
+def test_support_for_int():
+    """See issue: https://github.com/Lightning-AI/torchmetrics/issues/1970."""
+    metric = MulticlassStatScores(num_classes=4, average="none", multidim_average="samplewise", ignore_index=0)
+    prediction = torch.randint(low=0, high=4, size=(1, 224, 224)).to(torch.uint8)
+    label = torch.randint(low=0, high=4, size=(1, 224, 224)).to(torch.uint8)
+    score = metric(preds=prediction, target=label)
+    assert score.shape == (1, 4, 5)
+
+
 @pytest.mark.parametrize(
     ("metric", "kwargs"),
     [