BUG: NumericIndex should not support float16 dtype (#49536)

* BUG: NumericIndex should not support float16 dtype

* make NumericIndex fail with float16 dtype

* make NumericIndex fail with float16 dtype, II

* fix failures

* NotImplementedError

* NotImplementedError II

* fail on float16, but allow np.exp(int8_arrays)

* fail on float16, but allow np.exp(int8_arrays) II

* fix NumericIndex_ensure_dtype

Co-authored-by: Terji Petersen <[email protected]>
Co-authored-by: Terji Petersen <[email protected]>

Author: 3 people

pandas/core/algorithms.py

@@ -891,6 +891,8 @@ def value_counts(
         else:
             values = _ensure_arraylike(values)
             keys, counts = value_counts_arraylike(values, dropna)
+            if keys.dtype == np.float16:
+                keys = keys.astype(np.float32)
 
             # For backwards compatibility, we let Index do its normal type
             #  inference, _except_ for if if infers from object to bool.

pandas/core/base.py

@@ -1140,6 +1140,8 @@ def factorize(
         codes, uniques = algorithms.factorize(
             self._values, sort=sort, use_na_sentinel=use_na_sentinel
         )
+        if uniques.dtype == np.float16:
+            uniques = uniques.astype(np.float32)
 
         if isinstance(self, ABCIndex):
             # preserve e.g. NumericIndex, preserve MultiIndex

pandas/core/indexes/base.py

@@ -881,6 +881,9 @@ def __array_ufunc__(self, ufunc: np.ufunc, method: str_t, *inputs, **kwargs):
             # i.e. np.divmod, np.modf, np.frexp
             return tuple(self.__array_wrap__(x) for x in result)
 
+        if result.dtype == np.float16:
+            result = result.astype(np.float32)
+
         return self.__array_wrap__(result)
 
     def __array_wrap__(self, result, context=None):

pandas/core/indexes/numeric.py

@@ -75,8 +75,8 @@ class NumericIndex(Index):
     Notes
     -----
     An NumericIndex instance can **only** contain numpy int64/32/16/8, uint64/32/16/8 or
-    float64/32/16 dtype. In particular, ``NumericIndex`` *can not* hold Pandas numeric
-    dtypes (:class:`Int64Dtype`, :class:`Int32Dtype` etc.).
+    float64/32 dtype. In particular, ``NumericIndex`` *can not* hold numpy float16
+    dtype or Pandas numeric dtypes (:class:`Int64Dtype`, :class:`Int32Dtype` etc.).
     """
 
     _typ = "numericindex"
@@ -133,6 +133,10 @@ def _ensure_array(cls, data, dtype, copy: bool):
         Ensure we have a valid array to pass to _simple_new.
         """
         cls._validate_dtype(dtype)
+        if dtype == np.float16:
+
+            # float16 not supported (no indexing engine)
+            raise NotImplementedError("float16 indexes are not supported")
 
         if not isinstance(data, (np.ndarray, Index)):
             # Coerce to ndarray if not already ndarray or Index
@@ -176,6 +180,10 @@ def _ensure_array(cls, data, dtype, copy: bool):
             raise ValueError("Index data must be 1-dimensional")
 
         subarr = np.asarray(subarr)
+        if subarr.dtype == "float16":
+            # float16 not supported (no indexing engine)
+            raise NotImplementedError("float16 indexes are not implemented")
+
         return subarr
 
     @classmethod
@@ -202,6 +210,9 @@ def _ensure_dtype(cls, dtype: Dtype | None) -> np.dtype | None:
         dtype = pandas_dtype(dtype)
         if not isinstance(dtype, np.dtype):
             raise TypeError(f"{dtype} not a numpy type")
+        elif dtype == np.float16:
+            # float16 not supported (no indexing engine)
+            raise NotImplementedError("float16 indexes are not supported")
 
         if cls._is_backward_compat_public_numeric_index:
             # dtype for NumericIndex

pandas/tests/arithmetic/test_numeric.py

@@ -7,7 +7,6 @@
 from datetime import timedelta
 from decimal import Decimal
 import operator
-from typing import Any
 
 import numpy as np
 import pytest
@@ -72,15 +71,10 @@ def compare_op(series, other, op):
 # TODO: remove this kludge once mypy stops giving false positives here
 # List comprehension has incompatible type List[PandasObject]; expected List[RangeIndex]
 #  See GH#29725
-ser_or_index: list[Any] = [Series, Index]
-lefts: list[Any] = [RangeIndex(10, 40, 10)]
-lefts.extend(
-    [
-        cls([10, 20, 30], dtype=dtype)
-        for dtype in ["i1", "i2", "i4", "i8", "u1", "u2", "u4", "u8", "f2", "f4", "f8"]
-        for cls in ser_or_index
-    ]
-)
+_ldtypes = ["i1", "i2", "i4", "i8", "u1", "u2", "u4", "u8", "f2", "f4", "f8"]
+lefts: list[Index | Series] = [RangeIndex(10, 40, 10)]
+lefts.extend([Series([10, 20, 30], dtype=dtype) for dtype in _ldtypes])
+lefts.extend([Index([10, 20, 30], dtype=dtype) for dtype in _ldtypes if dtype != "f2"])
 
 # ------------------------------------------------------------------
 # Comparisons

pandas/tests/base/test_conversion.py

@@ -62,6 +62,10 @@ def test_iterable(self, index_or_series, method, dtype, rdtype):
         # gh-13258
         # coerce iteration to underlying python / pandas types
         typ = index_or_series
+        if dtype == "float16" and issubclass(typ, pd.Index):
+            with pytest.raises(NotImplementedError, match="float16 indexes are not "):
+                typ([1], dtype=dtype)
+            return
         s = typ([1], dtype=dtype)
         result = method(s)[0]
         assert isinstance(result, rdtype)
@@ -115,6 +119,10 @@ def test_iterable_map(self, index_or_series, dtype, rdtype):
         # gh-13236
         # coerce iteration to underlying python / pandas types
         typ = index_or_series
+        if dtype == "float16" and issubclass(typ, pd.Index):
+            with pytest.raises(NotImplementedError, match="float16 indexes are not "):
+                typ([1], dtype=dtype)
+            return
         s = typ([1], dtype=dtype)
         result = s.map(type)[0]
         if not isinstance(rdtype, tuple):

pandas/tests/base/test_value_counts.py

@@ -28,7 +28,13 @@ def test_value_counts(index_or_series_obj):
 
     counter = collections.Counter(obj)
     expected = Series(dict(counter.most_common()), dtype=np.int64, name=obj.name)
-    expected.index = expected.index.astype(obj.dtype)
+
+    if obj.dtype != np.float16:
+        expected.index = expected.index.astype(obj.dtype)
+    else:
+        with pytest.raises(NotImplementedError, match="float16 indexes are not "):
+            expected.index.astype(obj.dtype)
+        return
 
     if not isinstance(result.dtype, np.dtype):
         # i.e IntegerDtype
@@ -73,7 +79,13 @@ def test_value_counts_null(null_obj, index_or_series_obj):
     # np.nan would be duplicated, whereas None wouldn't
     counter = collections.Counter(obj.dropna())
     expected = Series(dict(counter.most_common()), dtype=np.int64)
-    expected.index = expected.index.astype(obj.dtype)
+
+    if obj.dtype != np.float16:
+        expected.index = expected.index.astype(obj.dtype)
+    else:
+        with pytest.raises(NotImplementedError, match="float16 indexes are not "):
+            expected.index.astype(obj.dtype)
+        return
 
     result = obj.value_counts()
     if obj.duplicated().any():

pandas/tests/indexes/numeric/test_numeric.py

@@ -471,6 +471,44 @@ def test_coerce_list(self):
         assert type(arr) is Index
 
 
+class TestFloat16Index:
+    # float 16 indexes not supported
+    # GH 49535
+    _index_cls = NumericIndex
+
+    def test_constructor(self):
+        index_cls = self._index_cls
+        dtype = np.float16
+
+        msg = "float16 indexes are not supported"
+
+        # explicit construction
+        with pytest.raises(NotImplementedError, match=msg):
+            index_cls([1, 2, 3, 4, 5], dtype=dtype)
+
+        with pytest.raises(NotImplementedError, match=msg):
+            index_cls(np.array([1, 2, 3, 4, 5]), dtype=dtype)
+
+        with pytest.raises(NotImplementedError, match=msg):
+            index_cls([1.0, 2, 3, 4, 5], dtype=dtype)
+
+        with pytest.raises(NotImplementedError, match=msg):
+            index_cls(np.array([1.0, 2, 3, 4, 5]), dtype=dtype)
+
+        with pytest.raises(NotImplementedError, match=msg):
+            index_cls([1.0, 2, 3, 4, 5], dtype=dtype)
+
+        with pytest.raises(NotImplementedError, match=msg):
+            index_cls(np.array([1.0, 2, 3, 4, 5]), dtype=dtype)
+
+        # nan handling
+        with pytest.raises(NotImplementedError, match=msg):
+            index_cls([np.nan, np.nan], dtype=dtype)
+
+        with pytest.raises(NotImplementedError, match=msg):
+            index_cls(np.array([np.nan]), dtype=dtype)
+
+
 class TestUIntNumericIndex(NumericInt):
 
     _index_cls = NumericIndex

pandas/tests/indexes/test_numpy_compat.py

@@ -77,7 +77,10 @@ def test_numpy_ufuncs_basic(index, func):
         # coerces to float (e.g. np.sin)
         with np.errstate(all="ignore"):
             result = func(index)
-            exp = Index(func(index.values), name=index.name)
+            arr_result = func(index.values)
+            if arr_result.dtype == np.float16:
+                arr_result = arr_result.astype(np.float32)
+            exp = Index(arr_result, name=index.name)
 
         tm.assert_index_equal(result, exp)
         if type(index) is not Index or index.dtype == bool:

pandas/tests/test_algos.py

@@ -66,7 +66,10 @@ def test_factorize(self, index_or_series_obj, sort):
         constructor = Index
         if isinstance(obj, MultiIndex):
             constructor = MultiIndex.from_tuples
-        expected_uniques = constructor(obj.unique())
+        expected_arr = obj.unique()
+        if expected_arr.dtype == np.float16:
+            expected_arr = expected_arr.astype(np.float32)
+        expected_uniques = constructor(expected_arr)
         if (
             isinstance(obj, Index)
             and expected_uniques.dtype == bool