testing.assert_close

array_api_extra.testing.assert_close(actual, desired, *, rtol=None, atol=0, equal_nan=True, err_msg='', verbose=True, check_dtype=True, check_shape=True, check_scalar=False, xp=None)

Check that two arrays are close, up to tolerance atol + rtol * abs(desired).

This is an interface to numpy.testing.assert_allclose() which accepts any standard-compatible array and performs additional array namespace, shape, and dtype checks.

Parameters:

  • actual (Array) – The array produced by the tested function.

  • desired (Array) – The expected array (typically hardcoded).

  • rtol (float or Array, optional) – Relative tolerance. Default: dtype-dependent.

  • atol (float or Array, optional) – Absolute tolerance. Default: 0.

  • equal_nan (bool, default: True) – Whether to consider NaNs in corresponding locations as equal.

  • err_msg (str, optional) – Error message to display on failure.

  • verbose (bool, default: True) – Whether to include the conflicting arrays in the error message on failure.

  • check_dtype (bool, default: True) – Whether to check agreement between actual and desired dtypes.

  • check_shape (bool, default: True) – Whether to check agreement between actual and desired shapes.

  • check_scalar (bool, default: False) – NumPy only: whether to check agreement between actual and desired types — 0-D numpy.ndarray vs scalar (e.g. numpy.double).

  • xp (array_namespace, optional) – A standard-compatible namespace which actual and desired must match.

Raises:

ImportError – If numpy is not importable in the Python environment.

Return type:

None

See also

assert_equal

Similar function for exact equality checks.

array_api_extra.isclose

Similar function checking closeness, returning a bool.

numpy.testing.assert_allclose

Similar function for NumPy arrays.

Notes

The default atol and rtol differ from xp.all(xpx.isclose(a, b)). For inexact dtypes, the default rtol is xp.finfo(actual.dtype).eps ** 0.5 * 4, which for float64 is roughly halfway between \(\sqrt{\epsilon}\) and the default for numpy.testing.assert_allclose(), 1e-7. This gives a more reasonable default for lower precision dtypes, for example approximately 1e-3 for float32. For exact dtypes, the default 1e-7 is used.

Array arguments to atol and rtol must be valid input to float.