Dtypes

[ev-br]

The current dtype stubs are too simplistic. Let's see if we can make them work without reproducing the whole story about dtypes [1] vs array scalars [2].
The minimum set of requirements is

• np.float32(42) should work. In numpy it creates a scalar, can we get around creating a single-element zero-dimensional array (backed by a zero-dimensional tensor).
• (a part of) dtype class hierarchy is needed. This includes common patterns like if np.issubdtype(x.dtype, np.complexfloating). Sadly, there are also isinstance checks in the numpy test suite:

(numpy-dev) br@gonzales:~/repos/numpy$ grep -rn np.inexact .
./doc/source/release/1.19.0-notes.rst:259:or ``np.inexact`` will now give a deprecation warning when converted
./doc/neps/nep-0040-legacy-datatype-impl.rst:254:types such as ``np.inexact`` (see figure below).
./doc/neps/nep-0040-legacy-datatype-impl.rst:256:``issubclass(a.dtype.type, np.inexact)``.
./doc/neps/nep-0042-new-dtypes.rst:330:   ``isinstance(np.dtype("float64"), np.inexact)``.
./numpy/random/bit_generator.pyx:139:    elif isinstance(x, (float, np.inexact)):
./numpy/linalg/tests/test_linalg.py:1229:        if issubclass(x.dtype.type, np.inexact):
./numpy/lib/function_base.py:1208:    elif np.issubdtype(otype, np.inexact):
./numpy/lib/function_base.py:3876:    if np.issubdtype(a.dtype, np.inexact):
./numpy/lib/function_base.py:3908:    if np.issubdtype(a.dtype, np.inexact) and sz > 0:
./numpy/lib/function_base.py:4681:    if np.issubdtype(arr.dtype, np.inexact):
./numpy/lib/function_base.py:4733:        if np.issubdtype(arr.dtype, np.inexact):
./numpy/lib/function_base.py:4753:        if np.issubdtype(arr.dtype, np.inexact):
./numpy/lib/nanfunctions.py:101:    elif issubclass(a.dtype.type, np.inexact):
./numpy/lib/nanfunctions.py:1039:    if dtype is not None and not issubclass(dtype.type, np.inexact):
./numpy/lib/nanfunctions.py:1041:    if out is not None and not issubclass(out.dtype.type, np.inexact):
./numpy/lib/nanfunctions.py:1720:    if dtype is not None and not issubclass(dtype.type, np.inexact):
./numpy/lib/nanfunctions.py:1722:    if out is not None and not issubclass(out.dtype.type, np.inexact):
./numpy/typing/tests/data/fail/scalars.pyi:56:np.inexact(1)  # E: Cannot instantiate abstract class
./numpy/ma/extras.py:742:    if np.issubdtype(a.dtype, np.inexact):
./numpy/ma/extras.py:772:        if np.issubdtype(asorted.dtype, np.inexact) and asorted.size > 0:
./numpy/ma/extras.py:812:    if np.issubdtype(asorted.dtype, np.inexact):
./numpy/core/tests/test_abc.py:12:        assert_(issubclass(np.inexact, numbers.Complex))
./numpy/core/tests/test_deprecations.py:426:        np.inexact, np.floating, np.complexfloating,
./numpy/core/tests/test_scalar_methods.py:140:        np.inexact,
./numpy/core/tests/test_numerictypes.py:381:        assert_(np.issubdtype(np.floating, np.inexact))
./numpy/core/tests/test_numerictypes.py:382:        assert_(not np.issubdtype(np.inexact, np.floating))
./numpy/core/tests/test_array_coercion.py:261:        if isinstance(scalar, np.inexact):

[1] https://numpy.org/doc/stable/reference/arrays.dtypes.html
[2] https://numpy.org/doc/stable/reference/arrays.scalars.html

[rgommers]

np.float32(42) should work. In numpy it creates a scalar, can we get around creating a single-element zero-dimensional array (backed by a zero-dimensional tensor).

If by "get around" you mean "avoid", then my question is why? At least for the "backed by a 0-D tensor" part I don't see another option.

[ev-br]

Meant "avoid having to distinguish between scalars and zero-dimensional arrays". See gh-13 for one attempt.