Added support for Structured Arrays

docs/changes.rst

@@ -82,5 +82,5 @@ Contributors
 ~~~~~~~~~~~~
 
 :user:`Adrin Jalali <adrinjalali>`, :user:`Merve Noyan <merveenoyan>`,
-:user:`Benjamin Bossan <BenjaminBossan>`, :user:`Ayyuce Demirbas
-<ayyucedemirbas>`, :user:`Prajjwal Mishra <p-mishra1>`
+:user:`Benjamin Bossan <BenjaminBossan>`, :user:`Ayyuce Demirbas <ayyucedemirbas>`,
+:user:`Prajjwal Mishra <p-mishra1>`, :user:`Ali Osman Kaya <aliosmankaya>`

skops/hub_utils/_hf_hub.py

@@ -132,7 +132,11 @@ def _get_column_names(data):
     # TODO: this is going to fail for Structured Arrays. We can add support for
     # them later if we see need for it.
     if isinstance(data, np.ndarray):
-        return [f"x{x}" for x in range(data.shape[1])]
+        if data.dtype.names:
+            return [f"x{x}" for x in range(len(data.dtype.names))]
+        return (
+            [f"x{x}" for x in range(data.shape[1])] if len(data.shape) > 1 else ["x0"]
+        )
 
     raise ValueError("The data is not a pandas.DataFrame or a numpy.ndarray.")
 
@@ -672,7 +676,18 @@ def get_model_output(repo_id: str, data: Any, token: Optional[str] = None) -> An
         inputs = {"data": data.to_dict(orient="list")}
     except AttributeError:
         # the input is not a pandas DataFrame
-        inputs = {f"x{i}": data[:, i] for i in range(data.shape[1])}
+        if data.dtype.names:
+            inputs = {
+                col: [x[0] for x in data[:, idx].tolist()]
+                if len(data.shape) > 1
+                else [x[0] for x in data.tolist()]
+                for idx, col in enumerate(_get_column_names(data))
+            }
+        else:
+            inputs = {
+                col: data[:, idx].tolist() if len(data.shape) > 1 else data.tolist()
+                for idx, col in enumerate(_get_column_names(data))
+            }
         inputs = {"data": inputs}
 
     res = InferenceApi(repo_id=repo_id, task=model_info.pipeline_tag, token=token)(

skops/hub_utils/tests/test_hf_hub.py

@@ -551,7 +551,18 @@ def test_get_column_names():
     expected_columns = [f"x{x}" for x in range(10)]
     assert _get_column_names(X_array) == expected_columns
 
-    expected_columns = [f"column{x}" for x in range(10)]
+    expected_columns = ["x0", "x1"]
+    X_array = np.array([(1, 2), (3, 4)], dtype=[("foo", "i8"), ("bar", "f4")])
+    assert _get_column_names(X_array) == expected_columns
+
+    expected_columns = ["x0", "x1", "x2"]  # Default names
+    X_array = np.zeros(3, dtype="int8, float32, float64")
+    assert _get_column_names(X_array) == expected_columns
+
+    expected_columns = ["x0"]
+    X_array = np.zeros(3)
+    assert _get_column_names(X_array) == expected_columns
+
     X_df = pd.DataFrame(X_array, columns=expected_columns)
     assert _get_column_names(X_df) == expected_columns