Fix/multiclass recall macro avg ignore index

CHANGELOG.md

@@ -53,7 +53,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
--
+- Fixed multiclass recall macro avg. ignore index ([#2710](https://github.com/Lightning-AI/torchmetrics/pull/2710))
+
+
+- Fixed iou scores in detection for either empty predictions/targets leading to wrong scores ([#2805](https://github.com/Lightning-AI/torchmetrics/pull/2805))
 
 
 ---

src/torchmetrics/classification/precision_recall.py

@@ -749,6 +749,7 @@ def compute(self) -> Tensor:
             multidim_average=self.multidim_average,
             top_k=self.top_k,
             zero_division=self.zero_division,
+            ignore_index=self.ignore_index,
         )
 
     def plot(

src/torchmetrics/functional/classification/accuracy.py

@@ -85,7 +85,9 @@ def _accuracy_reduce(
         return _safe_divide(tp, tp + fn)
 
     score = _safe_divide(tp + tn, tp + tn + fp + fn) if multilabel else _safe_divide(tp, tp + fn)
-    return _adjust_weights_safe_divide(score, average, multilabel, tp, fp, fn, top_k)
+    return _adjust_weights_safe_divide(
+        score=score, average=average, multilabel=multilabel, tp=tp, fp=fp, fn=fn, top_k=top_k
+    )
 
 
 def binary_accuracy(

src/torchmetrics/functional/classification/f_beta.py

@@ -55,7 +55,7 @@ def _fbeta_reduce(
         return _safe_divide((1 + beta2) * tp, (1 + beta2) * tp + beta2 * fn + fp, zero_division)
 
     fbeta_score = _safe_divide((1 + beta2) * tp, (1 + beta2) * tp + beta2 * fn + fp, zero_division)
-    return _adjust_weights_safe_divide(fbeta_score, average, multilabel, tp, fp, fn)
+    return _adjust_weights_safe_divide(score=fbeta_score, average=average, multilabel=multilabel, tp=tp, fp=fp, fn=fn)
 
 
 def _binary_fbeta_score_arg_validation(

src/torchmetrics/functional/classification/hamming.py

@@ -80,7 +80,7 @@ def _hamming_distance_reduce(
         return 1 - _safe_divide(tp, tp + fn)
 
     score = 1 - _safe_divide(tp + tn, tp + tn + fp + fn) if multilabel else 1 - _safe_divide(tp, tp + fn)
-    return _adjust_weights_safe_divide(score, average, multilabel, tp, fp, fn)
+    return _adjust_weights_safe_divide(score=score, average=average, multilabel=multilabel, tp=tp, fp=fp, fn=fn)
 
 
 def binary_hamming_distance(

src/torchmetrics/functional/classification/negative_predictive_value.py

@@ -53,7 +53,9 @@ def _negative_predictive_value_reduce(
         fn = fn.sum(dim=0 if multidim_average == "global" else 1)
         return _safe_divide(tn, tn + fn, zero_division)
     score = _safe_divide(tn, tn + fn, zero_division)
-    return _adjust_weights_safe_divide(score, average, multilabel, tp, fp, fn, top_k=top_k)
+    return _adjust_weights_safe_divide(
+        score=score, average=average, multilabel=multilabel, tp=tp, fp=fp, fn=fn, top_k=top_k
+    )
 
 
 def binary_negative_predictive_value(

src/torchmetrics/functional/classification/precision_recall.py

@@ -45,6 +45,7 @@ def _precision_recall_reduce(
     multilabel: bool = False,
     top_k: int = 1,
     zero_division: float = 0,
+    ignore_index: Optional[int] = None,
 ) -> Tensor:
     different_stat = fp if stat == "precision" else fn  # this is what differs between the two scores
     if average == "binary":
@@ -56,7 +57,7 @@ def _precision_recall_reduce(
         return _safe_divide(tp, tp + different_stat, zero_division)
 
     score = _safe_divide(tp, tp + different_stat, zero_division)
-    return _adjust_weights_safe_divide(score, average, multilabel, tp, fp, fn, top_k=top_k)
+    return _adjust_weights_safe_divide(score, average, multilabel, tp, fp, fn, top_k=top_k, ignore_index=ignore_index)
 
 
 def binary_precision(

src/torchmetrics/functional/classification/specificity.py

@@ -51,7 +51,9 @@ def _specificity_reduce(
         return _safe_divide(tn, tn + fp)
 
     specificity_score = _safe_divide(tn, tn + fp)
-    return _adjust_weights_safe_divide(specificity_score, average, multilabel, tp, fp, fn)
+    return _adjust_weights_safe_divide(
+        score=specificity_score, average=average, multilabel=multilabel, tp=tp, fp=fp, fn=fn
+    )
 
 
 def binary_specificity(

src/torchmetrics/utilities/compute.py

@@ -68,7 +68,14 @@ def _safe_divide(num: Tensor, denom: Tensor, zero_division: float = 0.0) -> Tens
 
 
 def _adjust_weights_safe_divide(
-    score: Tensor, average: Optional[str], multilabel: bool, tp: Tensor, fp: Tensor, fn: Tensor, top_k: int = 1
+    score: Tensor,
+    average: Optional[str],
+    multilabel: bool,
+    tp: Tensor,
+    fp: Tensor,
+    fn: Tensor,
+    top_k: int = 1,
+    ignore_index: Optional[int] = None,
 ) -> Tensor:
     if average is None or average == "none":
         return score
@@ -78,6 +85,11 @@ def _adjust_weights_safe_divide(
         weights = torch.ones_like(score)
         if not multilabel:
             weights[tp + fp + fn == 0 if top_k == 1 else tp + fn == 0] = 0.0
+
+    # Add this line to handle ignore_index
+    if ignore_index is not None and 0 <= ignore_index < len(score):
+        weights[ignore_index] = 0.0
+
     return _safe_divide(weights * score, weights.sum(-1, keepdim=True)).sum(-1)
 
 

tests/unittests/classification/test_precision_recall.py

@@ -661,6 +661,34 @@ def test_corner_case():
     assert res == 1.0
 
 
+@pytest.mark.parametrize(
+    ("ignore_index", "average", "expected"),
+    [
+        (0, "macro", 0.5),
+        (1, "macro", 1.0),
+        (None, "macro", 0.75),
+        (0, "none", torch.tensor([0.0, 0.5])),
+    ],
+)
+def test_multiclass_recall_ignore_index(ignore_index, average, expected):
+    """Issue: https://github.com/Lightning-AI/torchmetrics/issues/2441."""
+    y_true = torch.tensor([0, 0, 1, 1])
+    y_pred = torch.tensor([
+        [0.9, 0.1],
+        [0.9, 0.1],
+        [0.9, 0.1],
+        [0.1, 0.9],
+    ])
+
+    # Test with ignore_index=0 and average="macro"
+    metric_ignore = MulticlassRecall(num_classes=2, ignore_index=ignore_index, average=average)
+    res_ignore = metric_ignore(y_pred, y_true)
+    if isinstance(expected, float):
+        assert res_ignore == expected, f"Expected {expected}, but got {res_ignore}"
+    else:
+        assert torch.allclose(res_ignore, expected), f"Expected {expected}, but got {res_ignore}"
+
+
 @pytest.mark.parametrize(
     ("metric", "kwargs", "base_metric"),
     [