[YDF] Emit errors when using lists for multi-dim features

For NUMERICAL and BOOLEAN features, an incorrect dataset would be consumed before this change.

PiperOrigin-RevId: 691391034

yggdrasil_decision_forests/port/python/ydf/dataset/dataset.py

@@ -77,6 +77,13 @@ def normalize_categorical_string_value(value: Any) -> bytes:
             " task is selected. For example, you cannot train a classification"
             " model (task=ydf.Task.CLASSIFICATION) with floating point labels."
         )
+      if isinstance(value, list):
+        raise ValueError(
+            f"Cannot import column {column.name!r} with"
+            f" semantic={column.semantic} as it contains lists.\nNote:"
+            " Unrolling multi-dimensional columns is only supported for numpy"
+            " arrays"
+        )
       raise ValueError(
           f"Cannot import column {column.name!r} with"
           f" semantic={column.semantic} and"
@@ -134,6 +141,13 @@ def _add_column(
               " a regression model (task=ydf.Task.REGRESSION) on a string"
               " column."
           ) from e
+      if column_data.ndim != 1:
+        raise ValueError(
+            f"Cannot convert {column.semantic.name} column {column.name!r} "
+            f" with content={column_data!r} to a 1-dimensional array of"
+            " np.float32 values. Note: Unrolling multi-dimensional columns is"
+            " only supported for numpy arrays"
+        )
 
       if column.semantic == dataspec.Semantic.NUMERICAL:
         self._dataset.PopulateColumnNumericalNPFloat32(
@@ -172,6 +186,13 @@ def _add_column(
             f" values. Got {original_column_data!r}."
         )
         raise ValueError(message)
+      if column_data.ndim != 1:
+        raise ValueError(
+            f"Cannot convert BOOLEAN column {column.name!r}"
+            f" with content={column_data!r} to a 1-dimensional array of"
+            " np.float32 values. Note: Unrolling multi-dimensional columns is"
+            " only supported for numpy arrays"
+        )
 
       self._dataset.PopulateColumnBooleanNPBool(
           column.name,
@@ -221,6 +242,7 @@ def _add_column(
           )
         message += f"\nGot {original_column_data!r}."
         raise ValueError(message)
+      assert column_data.ndim == 1, "Categorical columns must be 1-dimensional"
 
       if column_data.dtype.type == np.bytes_:
         if inference_args is not None:

yggdrasil_decision_forests/port/python/ydf/dataset/dataset_test.py

@@ -1561,6 +1561,45 @@ def test_fail_gracefully_for_incorrect_boolean_type(self):
           columns=[("f1", dataspec.Semantic.BOOLEAN)],
       )
 
+  def test_multidim_numerical_list(self):
+    data = {
+        "f1": [[1, 2], [3, 4], [5, 6]],
+    }
+    with self.assertRaisesRegex(
+        ValueError,
+        ".*Unrolling multi-dimensional columns is only supported for numpy"
+        " arrays.*",
+    ):
+      _ = dataset.create_vertical_dataset(
+          data, columns=[("f1", dataspec.Semantic.NUMERICAL)]
+      )
+
+  def test_multidim_boolean_list(self):
+    data = {
+        "f1": [[True, False], [True, False], [True, False]],
+    }
+    with self.assertRaisesRegex(
+        ValueError,
+        ".*Unrolling multi-dimensional columns is only supported for numpy"
+        " arrays.*",
+    ):
+      _ = dataset.create_vertical_dataset(
+          data, columns=[("f1", dataspec.Semantic.BOOLEAN)]
+      )
+
+  def test_multidim_categorical_list(self):
+    data = {
+        "f1": [[1, 2], [3, 4], [5, 6]],
+    }
+    with self.assertRaisesRegex(
+        ValueError,
+        ".*Unrolling multi-dimensional columns is only supported for numpy"
+        " arrays.*",
+    ):
+      _ = dataset.create_vertical_dataset(
+          data, columns=[("f1", dataspec.Semantic.CATEGORICAL)]
+      )
+
 
 class CategoricalSetTest(absltest.TestCase):