Fix mixed types with cuDF. (#8280)

python-package/xgboost/core.py

@@ -1,49 +1,72 @@
 # pylint: disable=too-many-arguments, too-many-branches, invalid-name
 # pylint: disable=too-many-lines, too-many-locals
 """Core XGBoost Library."""
-from abc import ABC, abstractmethod
-from collections.abc import Mapping
 import copy
-from typing import List, Optional, Any, Union, Dict, TypeVar
-from typing import Callable, Tuple, cast, Sequence, Type, Iterable
 import ctypes
+import json
 import os
 import re
 import sys
-import json
 import warnings
+from abc import ABC, abstractmethod
+from collections.abc import Mapping
 from functools import wraps
-from inspect import signature, Parameter
+from inspect import Parameter, signature
+from typing import (
+    Any,
+    Callable,
+    Dict,
+    Iterable,
+    List,
+    Optional,
+    Sequence,
+    Tuple,
+    Type,
+    TypeVar,
+    Union,
+    cast,
+    overload,
+)
 
 import numpy as np
 import scipy.sparse
 
-from .compat import DataFrame, py_str, PANDAS_INSTALLED
-from .libpath import find_lib_path
 from ._typing import (
-    CStrPptr,
+    _T,
-    c_bst_ulong,
+    ArrayLike,
+    BoosterParam,
+    CFloatPtr,
     CNumeric,
-    DataType,
     CNumericPtr,
+    CStrPptr,
     CStrPtr,
     CTypeT,
-    ArrayLike,
-    CFloatPtr,
-    NumpyOrCupy,
-    FeatureInfo,
-    FeatureTypes,
-    FeatureNames,
-    _T,
     CupyT,
-    BoosterParam
+    DataType,
+    FeatureInfo,
+    FeatureNames,
+    FeatureTypes,
+    NumpyOrCupy,
+    c_bst_ulong,
 )
+from .compat import PANDAS_INSTALLED, DataFrame, py_str
+from .libpath import find_lib_path
 
 
 class XGBoostError(ValueError):
     """Error thrown by xgboost trainer."""
 
 
+@overload
+def from_pystr_to_cstr(data: str) -> bytes:
+    ...
+
+
+@overload
+def from_pystr_to_cstr(data: List[str]) -> ctypes.Array:
+    ...
+
+
 def from_pystr_to_cstr(data: Union[str, List[str]]) -> Union[bytes, ctypes.Array]:
     """Convert a Python str or list of Python str to C pointer
 

python-package/xgboost/data.py

@@ -3,24 +3,33 @@
 '''Data dispatching for DMatrix.'''
 import ctypes
 import json
-import warnings
 import os
-from typing import Any, Tuple, Callable, Optional, List, Union, Iterator, Sequence, cast
+import warnings
+from typing import Any, Callable, Iterator, List, Optional, Sequence, Tuple, Union, cast
 
 import numpy as np
 
-from .core import c_array, _LIB, _check_call, c_str
-from .core import _cuda_array_interface
-from .core import DataIter, _ProxyDMatrix, DMatrix
-from .compat import lazy_isinstance, DataFrame
 from ._typing import (
-    c_bst_ulong,
-    DataType,
-    FeatureTypes,
-    FeatureNames,
-    NumpyDType,
     CupyT,
-    FloatCompatible, PandasDType
+    DataType,
+    FeatureNames,
+    FeatureTypes,
+    FloatCompatible,
+    NumpyDType,
+    PandasDType,
+    c_bst_ulong,
+)
+from .compat import DataFrame, lazy_isinstance
+from .core import (
+    _LIB,
+    DataIter,
+    DMatrix,
+    _check_call,
+    _cuda_array_interface,
+    _ProxyDMatrix,
+    c_array,
+    c_str,
+    from_pystr_to_cstr,
 )
 
 DispatchedDataBackendReturnType = Tuple[
@@ -631,10 +640,10 @@ def _is_cudf_df(data: DataType) -> bool:
 
 
 def _cudf_array_interfaces(data: DataType, cat_codes: list) -> bytes:
-    """Extract CuDF __cuda_array_interface__.  This is special as it returns a new list of
+    """Extract CuDF __cuda_array_interface__.  This is special as it returns a new list
-    data and a list of array interfaces.  The data is list of categorical codes that
+    of data and a list of array interfaces.  The data is list of categorical codes that
-    caller can safely ignore, but have to keep their reference alive until usage of array
+    caller can safely ignore, but have to keep their reference alive until usage of
-    interface is finished.
+    array interface is finished.
 
     """
     try:
@@ -643,14 +652,18 @@ def _cudf_array_interfaces(data: DataType, cat_codes: list) -> bytes:
         from cudf.utils.dtypes import is_categorical_dtype
 
     interfaces = []
+
+    def append(interface: dict) -> None:
+        if "mask" in interface:
+            interface["mask"] = interface["mask"].__cuda_array_interface__
+        interfaces.append(interface)
+
     if _is_cudf_ser(data):
         if is_categorical_dtype(data.dtype):
             interface = cat_codes[0].__cuda_array_interface__
         else:
             interface = data.__cuda_array_interface__
-        if "mask" in interface:
+        append(interface)
-            interface["mask"] = interface["mask"].__cuda_array_interface__
-        interfaces.append(interface)
     else:
         for i, col in enumerate(data):
             if is_categorical_dtype(data[col].dtype):
@@ -658,10 +671,8 @@ def _cudf_array_interfaces(data: DataType, cat_codes: list) -> bytes:
                 interface = codes.__cuda_array_interface__
             else:
                 interface = data[col].__cuda_array_interface__
-            if "mask" in interface:
+            append(interface)
-                interface["mask"] = interface["mask"].__cuda_array_interface__
+    interfaces_str = from_pystr_to_cstr(json.dumps(interfaces))
-            interfaces.append(interface)
-    interfaces_str = bytes(json.dumps(interfaces, indent=2), "utf-8")
     return interfaces_str
 
 
@@ -722,9 +733,14 @@ def _transform_cudf_df(
             cat_codes.append(codes)
     else:
         for col in data:
-            if is_categorical_dtype(data[col].dtype) and enable_categorical:
+            dtype = data[col].dtype
+            if is_categorical_dtype(dtype) and enable_categorical:
                 codes = data[col].cat.codes
                 cat_codes.append(codes)
+            elif is_categorical_dtype(dtype):
+                raise ValueError(_ENABLE_CAT_ERR)
+            else:
+                cat_codes.append([])
 
     return data, cat_codes, feature_names, feature_types
 

tests/python-gpu/test_from_cudf.py

@@ -1,6 +1,8 @@
+import json
+import sys
+
 import numpy as np
 import xgboost as xgb
-import sys
 import pytest
 
 sys.path.append("tests/python")
@@ -176,20 +178,38 @@ Arrow specification.'''
         _test_cudf_metainfo(xgb.DeviceQuantileDMatrix)
 
     @pytest.mark.skipif(**tm.no_cudf())
-    def test_cudf_categorical(self):
+    def test_cudf_categorical(self) -> None:
         import cudf
-        _X, _y = tm.make_categorical(100, 30, 17, False)
+        n_features = 30
+        _X, _y = tm.make_categorical(100, n_features, 17, False)
         X = cudf.from_pandas(_X)
         y = cudf.from_pandas(_y)
 
         Xy = xgb.DMatrix(X, y, enable_categorical=True)
+        assert Xy.feature_types is not None
         assert len(Xy.feature_types) == X.shape[1]
         assert all(t == "c" for t in Xy.feature_types)
 
         Xy = xgb.DeviceQuantileDMatrix(X, y, enable_categorical=True)
+        assert Xy.feature_types is not None
         assert len(Xy.feature_types) == X.shape[1]
         assert all(t == "c" for t in Xy.feature_types)
 
+        # mixed dtypes
+        X["1"] = X["1"].astype(np.int64)
+        X["3"] = X["3"].astype(np.int64)
+        df, cat_codes, _, _ = xgb.data._transform_cudf_df(
+            X, None, None, enable_categorical=True
+        )
+        assert X.shape[1] == n_features
+        assert len(cat_codes) == X.shape[1]
+        assert not cat_codes[0]
+        assert not cat_codes[2]
+
+        interfaces_str = xgb.data._cudf_array_interfaces(df, cat_codes)
+        interfaces = json.loads(interfaces_str)
+        assert len(interfaces) == X.shape[1]
+
         # test missing value
         X = cudf.DataFrame({"f0": ["a", "b", np.NaN]})
         X["f0"] = X["f0"].astype("category")
@@ -206,7 +226,7 @@ Arrow specification.'''
         assert Xy.num_row() == 3
         assert Xy.num_col() == 1
 
-        with pytest.raises(ValueError):
+        with pytest.raises(ValueError, match="enable_categorical"):
             xgb.DeviceQuantileDMatrix(X, y)
 
         Xy = xgb.DeviceQuantileDMatrix(X, y, enable_categorical=True)