Apply new input validation to metrics.regression

python/cuml/cuml/metrics/regression.py

@@ -1,11 +1,17 @@
 #
-# SPDX-FileCopyrightText: Copyright (c) 2019-2025, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2019-2026, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 #
 import cupy as cp
 import numpy as np
 
-from cuml.internals.input_utils import input_to_cupy_array
+from cuml.internals.validation import (
+    check_array,
+    check_consistent_length,
+    check_sample_weight,
+)
+
+_FLOAT_OR_INT = (np.float32, np.float64, np.int32, np.int64)
 
 
 def _normalize_regression_metric_args(
@@ -16,31 +22,39 @@ def _normalize_regression_metric_args(
 
     Validates inputs and coerces all arrays to cupy of proper shape and dtype.
     """
-    # Coerce inputs to cupy arrays
-    float_or_int = [np.float32, np.float64, np.int32, np.int64]
-    y_true, n_rows, n_cols, _ = input_to_cupy_array(
-        y_true, check_dtype=float_or_int
+    y_true = check_array(
+        y_true, ensure_2d=False, dtype=_FLOAT_OR_INT, input_name="y_true"
     )
-    y_pred, _, _, _ = input_to_cupy_array(
-        y_pred, check_dtype=float_or_int, check_rows=n_rows, check_cols=n_cols
+    y_pred = check_array(
+        y_pred, ensure_2d=False, dtype=_FLOAT_OR_INT, input_name="y_pred"
     )
-    if sample_weight is not None:
-        sample_weight, _, _, _ = input_to_cupy_array(
-            sample_weight,
-            check_dtype=float_or_int,
-            check_rows=n_rows,
-            check_cols=1,
+    check_consistent_length(y_true, y_pred)
+
+    # Treat (N,) and (N, 1) as equivalent (matches sklearn's behavior in
+    # `_check_reg_targets`). Only reject genuine multi-output mismatches.
+    if y_true.ndim == 2 and y_true.shape[1] == 1:
+        y_true = y_true.ravel()
+    if y_pred.ndim == 2 and y_pred.shape[1] == 1:
+        y_pred = y_pred.ravel()
+
+    if y_true.ndim != y_pred.ndim or (
+        y_true.ndim == 2 and y_true.shape[1] != y_pred.shape[1]
+    ):
+        raise ValueError(
+            f"y_true and y_pred have different shapes: "
+            f"{y_true.shape} vs {y_pred.shape}"
         )
 
-    # Ensure y_true & y_pred are 2D and sample_weight is 1D
+    if (
+        sample_weight := check_sample_weight(sample_weight, dtype=np.float64)
+    ) is not None:
+        check_consistent_length(y_true, sample_weight)
+
+    # Promote 1D inputs to column vectors
     if y_true.ndim == 1:
         y_true = y_true.reshape((-1, 1))
-
-    if y_pred.ndim == 1:
         y_pred = y_pred.reshape((-1, 1))
-
-    if sample_weight is not None:
-        sample_weight = sample_weight.reshape(-1)
+    n_cols = y_true.shape[1]
 
     # Validate multioutput, and maybe coerce to a cupy array
     valid_multioutput = ("raw_values", "uniform_average", "variance_weighted")
@@ -54,9 +68,13 @@ def _normalize_regression_metric_args(
             raise ValueError(
                 "Custom weights are useful only in multi-output cases."
             )
-        multioutput, _, _, _ = input_to_cupy_array(
-            multioutput, check_rows=n_cols
+        multioutput = check_array(
+            multioutput, ensure_2d=False, input_name="multioutput"
         )
+        if multioutput.ndim != 1 or multioutput.shape[0] != n_cols:
+            raise ValueError(
+                f"There must be equally many custom weights ({n_cols}) as outputs."
+            )
 
     return y_true, y_pred, sample_weight, multioutput
 

python/cuml/cuml_accel_tests/upstream/scikit-learn/xfail-list.yaml

@@ -462,12 +462,10 @@
   - "sklearn.tests.test_common::test_estimators[SpectralCoclustering()-check_fit2d_1sample]"
   - "sklearn.tests.test_common::test_estimators[SpectralCoclustering()-check_methods_subset_invariance]"
   - "sklearn.tests.test_common::test_search_cv[GridSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('logisticregression',LogisticRegression())]),param_grid={'logisticregression__C':[0.1,1.0]})-check_fit2d_1feature]"
-  - "sklearn.tests.test_common::test_search_cv[GridSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_grid={'ridge__alpha':[0.1,1.0]})-check_dtype_object]"
   - "sklearn.tests.test_common::test_search_cv[GridSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_grid={'ridge__alpha':[0.1,1.0]})-check_fit2d_1feature]"
   - "sklearn.tests.test_common::test_search_cv[GridSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_grid={'ridge__alpha':[0.1,1.0]})-check_supervised_y_2d]"
   - "sklearn.tests.test_common::test_search_cv[GridSearchCV(cv=2,estimator=Ridge(),param_grid={'alpha':[0.1,1.0]})-check_supervised_y_2d]"
   - "sklearn.tests.test_common::test_search_cv[HalvingGridSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('logisticregression',LogisticRegression())]),min_resources='smallest',param_grid={'logisticregression__C':[0.1,1.0]},random_state=0)-check_fit2d_1feature]"
-  - "sklearn.tests.test_common::test_search_cv[HalvingGridSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),min_resources='smallest',param_grid={'ridge__alpha':[0.1,1.0]},random_state=0)-check_dtype_object]"
   - "sklearn.tests.test_common::test_search_cv[HalvingGridSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),min_resources='smallest',param_grid={'ridge__alpha':[0.1,1.0]},random_state=0)-check_estimators_fit_returns_self(readonly_memmap=True)]"
   - "sklearn.tests.test_common::test_search_cv[HalvingGridSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),min_resources='smallest',param_grid={'ridge__alpha':[0.1,1.0]},random_state=0)-check_estimators_fit_returns_self]"
   - "sklearn.tests.test_common::test_search_cv[HalvingGridSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),min_resources='smallest',param_grid={'ridge__alpha':[0.1,1.0]},random_state=0)-check_estimators_overwrite_params]"
@@ -480,7 +478,6 @@
   - "sklearn.tests.test_common::test_search_cv[HalvingGridSearchCV(cv=2,estimator=Ridge(),min_resources='smallest',param_grid={'alpha':[0.1,1.0]},random_state=0)-check_supervised_y_2d1]"
   - "sklearn.tests.test_common::test_search_cv[HalvingRandomSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('logisticregression',LogisticRegression())]),param_distributions={'logisticregression__C':[0.1,1.0]},random_state=0)-check_fit2d_1feature]"
   - "sklearn.tests.test_common::test_search_cv[HalvingRandomSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('logisticregression',LogisticRegression())]),param_distributions={'logisticregression__C':[0.1,1.0]},random_state=0)-check_fit2d_1sample]"
-  - "sklearn.tests.test_common::test_search_cv[HalvingRandomSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_distributions={'ridge__alpha':[0.1,1.0]},random_state=0)-check_dtype_object]"
   - "sklearn.tests.test_common::test_search_cv[HalvingRandomSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_distributions={'ridge__alpha':[0.1,1.0]},random_state=0)-check_estimators_fit_returns_self(readonly_memmap=True)]"
   - "sklearn.tests.test_common::test_search_cv[HalvingRandomSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_distributions={'ridge__alpha':[0.1,1.0]},random_state=0)-check_estimators_fit_returns_self]"
   - "sklearn.tests.test_common::test_search_cv[HalvingRandomSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_distributions={'ridge__alpha':[0.1,1.0]},random_state=0)-check_estimators_overwrite_params]"
@@ -489,7 +486,6 @@
   - "sklearn.tests.test_common::test_search_cv[HalvingRandomSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_distributions={'ridge__alpha':[0.1,1.0]},random_state=0)-check_supervised_y_2d]"
   - "sklearn.tests.test_common::test_search_cv[RandomizedSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('logisticregression',LogisticRegression())]),param_distributions={'logisticregression__C':[0.1,1.0]},random_state=0)-check_fit2d_1feature]"
   - "sklearn.tests.test_common::test_search_cv[RandomizedSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('logisticregression',LogisticRegression())]),param_distributions={'logisticregression__C':[0.1,1.0]},random_state=0)-check_supervised_y_2d]"
-  - "sklearn.tests.test_common::test_search_cv[RandomizedSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_distributions={'ridge__alpha':[0.1,1.0]},random_state=0)-check_dtype_object]"
   - "sklearn.tests.test_common::test_search_cv[RandomizedSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_distributions={'ridge__alpha':[0.1,1.0]},random_state=0)-check_fit2d_1feature]"
   - "sklearn.tests.test_common::test_search_cv[RandomizedSearchCV(cv=2,error_score='raise',estimator=Pipeline(steps=[('pca',PCA()),('ridge',Ridge())]),param_distributions={'ridge__alpha':[0.1,1.0]},random_state=0)-check_supervised_y_2d]"
   - "sklearn.tests.test_common::test_search_cv[RandomizedSearchCV(cv=2,estimator=Ridge(),param_distributions={'alpha':[0.1,1.0]},random_state=0)-check_supervised_y_2d]"

python/cuml/tests/test_metrics.py

@@ -16,6 +16,7 @@
 import sklearn.metrics
 from numba import cuda
 from numpy.testing import assert_almost_equal
+from packaging.version import Version
 from scipy.spatial import distance as scipy_pairwise_distances
 from scipy.special import rel_entr as scipy_kl_divergence
 from scipy.stats import entropy as sp_entropy
@@ -766,6 +767,102 @@ def test_mean_squared_log_error_negative_values(inputs):
         )
 
 
+_REGRESSION_FUNCS = [
+    "r2_score",
+    "mean_squared_error",
+    "mean_absolute_error",
+    "median_absolute_error",
+    "mean_squared_log_error",
+]
+
+
+@pytest.mark.parametrize("func", _REGRESSION_FUNCS)
+def test_regression_metrics_scalar_sample_weight(func):
+    y_true = np.array([1.0, 2.0, 3.0, 4.0])
+    y_pred = np.array([1.1, 1.9, 3.1, 3.9])
+
+    cu_metric = getattr(cuml.metrics, func)
+    skl_metric = getattr(sklearn.metrics, func)
+
+    unweighted = cu_metric(y_true, y_pred)
+    assert cu_metric(y_true, y_pred, sample_weight=1.0) == unweighted
+    assert cu_metric(y_true, y_pred, sample_weight=2.5) == unweighted
+
+    # Verify cuML matches scikit-learn for unweighted
+    np.testing.assert_allclose(
+        cu_metric(y_true, y_pred), skl_metric(y_true, y_pred)
+    )
+
+    # cuML accepts scalar sample_weight (equivalent to sample_weight=1.0).
+    # sklearn requires array-like sample_weight, so compare cuML scalar-sw
+    # with sklearn using a 1D array of ones.
+    sw_scalar = 1.0
+    sw_array = np.array([sw_scalar] * len(y_true))
+    np.testing.assert_allclose(
+        cu_metric(y_true, y_pred, sample_weight=sw_scalar),
+        skl_metric(y_true, y_pred, sample_weight=sw_array),
+    )
+
+
+@pytest.mark.parametrize("func", _REGRESSION_FUNCS)
+@pytest.mark.parametrize(
+    "y_true_shape, y_pred_shape",
+    [((4,), (4, 1)), ((4, 1), (4,)), ((4, 1), (4, 1))],
+)
+def test_regression_metrics_1d_2d_equivalence(
+    func, y_true_shape, y_pred_shape
+):
+    # (N,) and (N, 1) should be treated as equivalent (matches sklearn).
+    y_true_1d = np.array([1.0, 2.0, 3.0, 4.0])
+    y_pred_1d = np.array([1.1, 1.9, 3.1, 3.9])
+
+    cu_metric = getattr(cuml.metrics, func)
+    skl_metric = getattr(sklearn.metrics, func)
+    expected = skl_metric(y_true_1d, y_pred_1d)
+
+    got = cu_metric(
+        y_true_1d.reshape(y_true_shape), y_pred_1d.reshape(y_pred_shape)
+    )
+    np.testing.assert_allclose(got, expected)
+
+
+@pytest.mark.parametrize("func", _REGRESSION_FUNCS)
+@pytest.mark.xfail(
+    condition=Version(sklearn.__version__) < Version("1.7"),
+    reason=(
+        "sklearn < 1.7 uses different error messages for invalid "
+        "sample_weight inputs; messages were standardized in sklearn 1.7"
+    ),
+    strict=True,
+)
+def test_regression_metrics_errors(func):
+    arr_3 = np.array([1.0, 2.0, 3.0])
+    arr_4 = np.array([1.0, 2.0, 3.0, 4.0])
+    arr_3x2 = np.ones((3, 2))
+
+    cu_metric = getattr(cuml.metrics, func)
+    skl_metric = getattr(sklearn.metrics, func)
+
+    # cuML and sklearn both raise ValueError for mismatched sample counts.
+    # sklearn: "Found input variables with inconsistent numbers of samples"
+    # cuML: "inconsistent number of samples" — match common substring.
+    sw_mismatch = np.array([1.0, 2.0, 3.0, 4.0])
+    with pytest.raises(ValueError, match="inconsistent"):
+        skl_metric(arr_3, arr_4)
+    with pytest.raises(ValueError, match="inconsistent"):
+        cu_metric(arr_3, arr_4)
+
+    with pytest.raises(ValueError, match="inconsistent"):
+        skl_metric(arr_3, arr_3, sample_weight=sw_mismatch)
+    with pytest.raises(ValueError, match="inconsistent"):
+        cu_metric(arr_3, arr_3, sample_weight=sw_mismatch)
+
+    with pytest.raises(ValueError, match="Sample weights must be 1D"):
+        skl_metric(arr_3, arr_3, sample_weight=arr_3x2)
+    with pytest.raises(ValueError, match="Sample weights must be 1D"):
+        cu_metric(arr_3, arr_3, sample_weight=arr_3x2)
+
+
 def test_entropy():
     # The outcome of a fair coin is the most uncertain:
     # in base 2 the result is 1 (One bit of entropy).