Fix test failures under NumPy 1.24.

NumPy 1.24 release notes: https://numpy.org/devdocs/release/1.24.0-notes.html

The fixes vary, but there are three particularly common changes:
* NumPy 1.24 removes a number of deprecated NumPy type aliases references (np.bool, np.int, np.float, np.complex, np.object, np.str, np.unicode, np.long). This change replaces them with their recommended replacements (bool, int, float, complex, object, str, str, int).
* Under NumPy 1.24 no longer automatically infers dtype=object when ragged sequences are passed to np.array(). See https://numpy.org/neps/nep-0034-infer-dtype-is-object.html . In most cases the fix is to pass dtype=object explicitly, but in some cases where the raggedness seems accidental other fixes were used.
* NumPy 1.24 is pickier about the dtype= option passed to comparison ufuncs.

PiperOrigin-RevId: 502979152
This commit is contained in:
Peter Hawkins 2023-01-18 14:24:35 -08:00 committed by jax authors
parent d58266eac7
commit 1929d34cb6

View File

@ -158,7 +158,7 @@ JAX_ONE_TO_ONE_OP_RECORDS = [
op_record("arctan", 1, number_dtypes, all_shapes, jtu.rand_small, ["rev"],
inexact=True),
op_record("arctan2", 2, float_dtypes, all_shapes, jtu.rand_small, ["rev"],
inexact=True),
inexact=True, check_dtypes=False),
op_record("arcsinh", 1, number_dtypes, all_shapes, jtu.rand_default, ["rev"],
inexact=True, tolerance={np.complex64: 2E-4, np.complex128: 2E-14}),
op_record("arccosh", 1, number_dtypes, all_shapes, jtu.rand_default, ["rev"],