Fix compatibility with Pandas > 3.0

verde/base/utils.py

@@ -8,7 +8,7 @@
 Utility functions for building gridders and checking arguments.
 """
 import numpy as np
-from sklearn.base import BaseEstimator
+from sklearn.base import BaseEstimator, RegressorMixin
 from sklearn.metrics import check_scoring
 
 
@@ -51,7 +51,7 @@ def score_estimator(scoring, estimator, coordinates, data, weights=None):
         coordinates, data, weights, unpack=False
     )
     predicted = check_data(estimator.predict(coordinates))
-    scorer = check_scoring(DummyEstimator, scoring=scoring)
+    scorer = check_scoring(DummyEstimator(np.array([0])), scoring=scoring)
     result = np.mean(
         [
             scorer(
@@ -66,7 +66,7 @@ def score_estimator(scoring, estimator, coordinates, data, weights=None):
     return result
 
 
-class DummyEstimator(BaseEstimator):
+class DummyEstimator(RegressorMixin, BaseEstimator):
     """
     Dummy estimator that does nothing but pass along the predicted data.
     Used to fool the scikit-learn scorer functions to fit our API

verde/blockreduce.py

@@ -463,7 +463,7 @@ def _blocked_mean_variance(self, table, ncomps):
         The variance will be the unweighted variance of the blocks.
         """
         reduction = {
-            "data{}".format(i): (("mean", self.reduction), ("variance", np.var))
+            "data{}".format(i): (("mean", "mean"), ("variance", "var"))
             for i in range(ncomps)
         }
         blocked = table.groupby("block").aggregate(reduction)

verde/tests/test_model_selection.py

@@ -55,6 +55,15 @@ def test_cross_val_score_vector(trend, metric, expected):
     npt.assert_allclose(scores, expected, atol=1e-10)
 
 
+def test_gridder_score(trend):
+    "Check that gridder.score works with the default R² scorer."
+    coords, data = trend[:2]
+    model = Trend(degree=1).fit(coords, data)
+    with pytest.warns(FutureWarning):
+        score = model.score(coords, data)
+    npt.assert_allclose(score, 1, atol=1e-10)
+
+
 def test_cross_val_score_client(trend):
     "Test the deprecated dask Client interface"
     coords, data = trend[:2]

verde/tests/test_utils.py

@@ -11,6 +11,7 @@
 
 import numpy as np
 import numpy.testing as npt
+import pandas as pd
 import pytest
 import xarray as xr
 from scipy.spatial import cKDTree
@@ -28,6 +29,7 @@
     meshgrid_to_1d,
     parse_engine,
     partition_by_sum,
+    variance_to_weights,
 )
 
 
@@ -108,6 +110,21 @@ def test_partition_by_sum_fails_no_partitions():
     assert "Could not find partition points" in str(error)
 
 
+def test_variance_to_weights_pandas_series():
+    "A pandas Series input should work even if its backing array is read-only."
+    variance = pd.Series([0, 2, 0.2, np.nan])
+    weights = variance_to_weights(variance)
+    npt.assert_allclose(weights, [1, 0.1, 1, 1])
+
+
+def test_variance_to_weights_readonly_array():
+    "A read-only NumPy array input should still produce normalized weights."
+    variance = np.array([0, 2, 0.2, np.nan])
+    variance.flags.writeable = False
+    weights = variance_to_weights(variance)
+    npt.assert_allclose(weights, [1, 0.1, 1, 1])
+
+
 def test_make_xarray_grid():
     """
     Check if xarray.Dataset is correctly created

verde/utils.py

@@ -164,7 +164,9 @@ def variance_to_weights(variance, tol=1e-15, dtype="float64"):
     variance = check_data(variance)
     weights = []
     for var in variance:
-        var = np.nan_to_num(np.atleast_1d(var), copy=False)
+        # Pandas 3 can expose Series-backed arrays as read-only, so normalize
+        # NaNs on a private writeable copy.
+        var = np.nan_to_num(np.atleast_1d(var), copy=True)
         w = np.ones_like(var, dtype=dtype)
         nonzero = var > tol
         if np.any(nonzero):