fix: normalize category labels in functional nominal metrics for non-zero-based labels

src/torchmetrics/functional/nominal/cramers.py

@@ -25,6 +25,7 @@
     _drop_empty_rows_and_cols,
     _handle_nan_in_data,
     _nominal_input_validation,
+    _normalize_categorical_labels,
     _unable_to_use_bias_correction_warning,
 )
 
@@ -52,6 +53,7 @@ def _cramers_v_update(
     preds = preds.argmax(1) if preds.ndim == 2 else preds
     target = target.argmax(1) if target.ndim == 2 else target
     preds, target = _handle_nan_in_data(preds, target, nan_strategy, nan_replace_value)
+    preds, target = _normalize_categorical_labels(preds, target)
     return _multiclass_confusion_matrix_update(preds, target, num_classes)
 
 

src/torchmetrics/functional/nominal/pearson.py

@@ -24,6 +24,7 @@
     _drop_empty_rows_and_cols,
     _handle_nan_in_data,
     _nominal_input_validation,
+    _normalize_categorical_labels,
 )
 
 
@@ -50,6 +51,7 @@ def _pearsons_contingency_coefficient_update(
     preds = preds.argmax(1) if preds.ndim == 2 else preds
     target = target.argmax(1) if target.ndim == 2 else target
     preds, target = _handle_nan_in_data(preds, target, nan_strategy, nan_replace_value)
+    preds, target = _normalize_categorical_labels(preds, target)
     return _multiclass_confusion_matrix_update(preds, target, num_classes)
 
 

src/torchmetrics/functional/nominal/theils_u.py

@@ -23,6 +23,7 @@
     _drop_empty_rows_and_cols,
     _handle_nan_in_data,
     _nominal_input_validation,
+    _normalize_categorical_labels,
 )
 
 
@@ -75,6 +76,7 @@ def _theils_u_update(
     preds = preds.argmax(1) if preds.ndim == 2 else preds
     target = target.argmax(1) if target.ndim == 2 else target
     preds, target = _handle_nan_in_data(preds, target, nan_strategy, nan_replace_value)
+    preds, target = _normalize_categorical_labels(preds, target)
     return _multiclass_confusion_matrix_update(preds, target, num_classes)
 
 

src/torchmetrics/functional/nominal/tschuprows.py

@@ -25,6 +25,7 @@
     _drop_empty_rows_and_cols,
     _handle_nan_in_data,
     _nominal_input_validation,
+    _normalize_categorical_labels,
     _unable_to_use_bias_correction_warning,
 )
 
@@ -52,6 +53,7 @@ def _tschuprows_t_update(
     preds = preds.argmax(1) if preds.ndim == 2 else preds
     target = target.argmax(1) if target.ndim == 2 else target
     preds, target = _handle_nan_in_data(preds, target, nan_strategy, nan_replace_value)
+    preds, target = _normalize_categorical_labels(preds, target)
     return _multiclass_confusion_matrix_update(preds, target, num_classes)
 
 

src/torchmetrics/functional/nominal/utils.py

@@ -140,6 +140,28 @@ def _handle_nan_in_data(
     return preds[~rows_contain_nan], target[~rows_contain_nan]
 
 
+def _normalize_categorical_labels(preds: Tensor, target: Tensor) -> tuple[Tensor, Tensor]:
+    """Map observed labels to contiguous 0-based IDs.
+
+    Nominal association metrics are invariant to renaming categories, but the
+    confusion-matrix updater uses raw label values as bin indices. Labels like
+    [1, 2] (instead of [0, 1]) would create an out-of-range bin.
+
+    Returns:
+        preds and target relabeled as contiguous 0-based IDs, or the
+        originals when they are already contiguous from zero.
+
+    """
+    all_vals = torch.cat([preds, target])
+    unique_vals = all_vals.unique()
+    expected = torch.arange(len(unique_vals), device=unique_vals.device, dtype=unique_vals.dtype)
+    if torch.equal(unique_vals, expected):
+        return preds, target
+    preds = torch.searchsorted(unique_vals, preds)
+    target = torch.searchsorted(unique_vals, target)
+    return preds, target
+
+
 def _unable_to_use_bias_correction_warning(metric_name: str) -> None:
     rank_zero_warn(
         f"Unable to compute {metric_name} using bias correction. Please consider to set `bias_correction=False`."