Support half type for pandas. (#8481)

python-package/xgboost/data.py

@@ -30,6 +30,7 @@ from .core import (
     c_array,
     c_str,
     from_pystr_to_cstr,
+    make_jcargs,
 )
 
 DispatchedDataBackendReturnType = Tuple[
@@ -184,24 +185,15 @@ def _from_numpy_array(
     feature_names: Optional[FeatureNames],
     feature_types: Optional[FeatureTypes],
 ) -> DispatchedDataBackendReturnType:
-    """Initialize data from a 2-D numpy matrix.
-
-    """
+    """Initialize data from a 2-D numpy matrix."""
     if len(data.shape) != 2:
-        raise ValueError(
-            "Expecting 2 dimensional numpy.ndarray, got: ", data.shape
-        )
+        raise ValueError("Expecting 2 dimensional numpy.ndarray, got: ", data.shape)
     data, _ = _ensure_np_dtype(data, data.dtype)
     handle = ctypes.c_void_p()
-    args = {
-        "missing": float(missing),
-        "nthread": int(nthread),
-    }
-    config = bytes(json.dumps(args), "utf-8")
     _check_call(
         _LIB.XGDMatrixCreateFromDense(
             _array_interface(data),
-            config,
+            make_jcargs(missing=float(missing), nthread=int(nthread)),
             ctypes.byref(handle),
         )
     )
@@ -1205,6 +1197,7 @@ def _proxy_transform(
         arr, feature_names, feature_types = _transform_pandas_df(
             data, enable_categorical, feature_names, feature_types
         )
+        arr, _ = _ensure_np_dtype(arr, arr.dtype)
         return arr, None, feature_names, feature_types
     raise TypeError("Value type is not supported for data iterator:" + str(type(data)))
 

python-package/xgboost/testing/data.py

@@ -0,0 +1,62 @@
+"""Utilities for data generation."""
+from typing import Generator, Tuple
+
+import numpy as np
+
+
+def np_dtypes(
+    n_samples: int, n_features: int
+) -> Generator[Tuple[np.ndarray, np.ndarray], None, None]:
+    """Generate all supported dtypes from numpy."""
+    import pandas as pd
+
+    rng = np.random.RandomState(1994)
+    # Integer and float.
+    orig = rng.randint(low=0, high=127, size=n_samples * n_features).reshape(
+        n_samples, n_features
+    )
+    dtypes = [
+        np.int32,
+        np.int64,
+        np.byte,
+        np.short,
+        np.intc,
+        np.int_,
+        np.longlong,
+        np.uint32,
+        np.uint64,
+        np.ubyte,
+        np.ushort,
+        np.uintc,
+        np.uint,
+        np.ulonglong,
+        np.float16,
+        np.float32,
+        np.float64,
+        np.half,
+        np.single,
+        np.double,
+    ]
+    for dtype in dtypes:
+        X = np.array(orig, dtype=dtype)
+        yield orig, X
+
+    for dtype in dtypes:
+        X = np.array(orig, dtype=dtype)
+        df_orig = pd.DataFrame(orig)
+        df = pd.DataFrame(X)
+        yield df_orig, df
+
+    # Boolean
+    orig = rng.binomial(1, 0.5, size=n_samples * n_features).reshape(
+        n_samples, n_features
+    )
+    for dtype in [np.bool_, np.bool8, bool]:
+        X = np.array(orig, dtype=dtype)
+        yield orig, X
+
+    for dtype in [np.bool_, np.bool8, bool]:
+        X = np.array(orig, dtype=dtype)
+        df_orig = pd.DataFrame(orig)
+        df = pd.DataFrame(X)
+        yield df_orig, df

tests/python/test_dmatrix.py

@@ -6,6 +6,7 @@ import pytest
 import scipy.sparse
 from hypothesis import given, settings, strategies
 from scipy.sparse import csr_matrix, rand
+from xgboost.testing.data import np_dtypes
 
 import xgboost as xgb
 from xgboost import testing as tm
@@ -453,3 +454,15 @@ class TestDMatrix:
         np.testing.assert_equal(csr.indptr, ret.indptr)
         np.testing.assert_equal(csr.data, ret.data)
         np.testing.assert_equal(csr.indices, ret.indices)
+
+    def test_dtypes(self) -> None:
+        n_samples = 128
+        n_features = 16
+        for orig, x in np_dtypes(n_samples, n_features):
+            m0 = xgb.DMatrix(orig)
+            m1 = xgb.DMatrix(x)
+            csr0 = m0.get_data()
+            csr1 = m1.get_data()
+            np.testing.assert_allclose(csr0.data, csr1.data)
+            np.testing.assert_allclose(csr0.indptr, csr1.indptr)
+            np.testing.assert_allclose(csr0.indices, csr1.indices)

tests/python/test_predict.py

@@ -5,6 +5,7 @@ import numpy as np
 import pandas as pd
 import pytest
 from scipy import sparse
+from xgboost.testing.data import np_dtypes
 from xgboost.testing.shared import validate_leaf_output
 
 import xgboost as xgb
@@ -230,46 +231,10 @@ class TestInplacePredict:
         from_dmatrix = booster.predict(dtrain)
         np.testing.assert_allclose(from_dmatrix, from_inplace)
 
-    def test_dtypes(self):
-        orig = self.rng.randint(low=0, high=127, size=self.rows * self.cols).reshape(
-            self.rows, self.cols
-        )
-        predt_orig = self.booster.inplace_predict(orig)
-        # all primitive types in numpy
-        for dtype in [
-            np.int32,
-            np.int64,
-            np.byte,
-            np.short,
-            np.intc,
-            np.int_,
-            np.longlong,
-            np.uint32,
-            np.uint64,
-            np.ubyte,
-            np.ushort,
-            np.uintc,
-            np.uint,
-            np.ulonglong,
-            np.float16,
-            np.float32,
-            np.float64,
-            np.half,
-            np.single,
-            np.double,
-        ]:
-            X = np.array(orig, dtype=dtype)
-            predt = self.booster.inplace_predict(X)
-            np.testing.assert_allclose(predt, predt_orig)
-
-        # boolean
-        orig = self.rng.binomial(1, 0.5, size=self.rows * self.cols).reshape(
-            self.rows, self.cols
-        )
-        predt_orig = self.booster.inplace_predict(orig)
-        for dtype in [np.bool8, np.bool_]:
-            X = np.array(orig, dtype=dtype)
-            predt = self.booster.inplace_predict(X)
+    def test_dtypes(self) -> None:
+        for orig, x in np_dtypes(self.rows, self.cols):
+            predt_orig = self.booster.inplace_predict(orig)
+            predt = self.booster.inplace_predict(x)
             np.testing.assert_allclose(predt, predt_orig)
 
         # unsupported types
@@ -278,6 +243,6 @@ class TestInplacePredict:
             np.complex64,
             np.complex128,
         ]:
-            X = np.array(orig, dtype=dtype)
+            X: np.ndarray = np.array(orig, dtype=dtype)
             with pytest.raises(ValueError):
                 self.booster.inplace_predict(X)

tests/python/test_quantile_dmatrix.py

@@ -11,6 +11,7 @@ from xgboost.testing import (
     make_categorical,
     make_sparse_regression,
 )
+from xgboost.testing.data import np_dtypes
 
 import xgboost as xgb
 
@@ -238,3 +239,25 @@ class TestQuantileDMatrix:
         np.testing.assert_allclose(
             booster.predict(qdm), booster.predict(xgb.DMatrix(qdm.get_data()))
         )
+
+    def test_dtypes(self) -> None:
+        n_samples = 128
+        n_features = 16
+        for orig, x in np_dtypes(n_samples, n_features):
+            m0 = xgb.QuantileDMatrix(orig)
+            m1 = xgb.QuantileDMatrix(x)
+            csr0 = m0.get_data()
+            csr1 = m1.get_data()
+            np.testing.assert_allclose(csr0.data, csr1.data)
+            np.testing.assert_allclose(csr0.indptr, csr1.indptr)
+            np.testing.assert_allclose(csr0.indices, csr1.indices)
+
+        # unsupported types
+        for dtype in [
+            np.string_,
+            np.complex64,
+            np.complex128,
+        ]:
+            X: np.ndarray = np.array(orig, dtype=dtype)
+            with pytest.raises(ValueError):
+                xgb.QuantileDMatrix(X)