python_api: handle array-like args in approx()

This treats objects that expose an ndarray via the __array__ interface the
same as direct subclasses of ndarray. Fixes pytest-dev#8132.

changelog/8132.bugfix.rst

@@ -0,0 +1,10 @@
+Fixed regression in ``approx``: in 6.2.0 ``approx`` no longer raises
+``TypeError`` when dealing with non-numeric types, falling back to normal comparison.
+Before 6.2.0, array types like tf.DeviceArray fell through to the scalar case,
+and happened to compare correctly to a scalar if they had only one element.
+After 6.2.0, these types began failing, because they inherited neither from
+standard Python number hierarchy nor from ``numpy.ndarray``.
+
+``approx`` now converts arguments to ``numpy.ndarray`` if they expose the array
+protocol and are not scalars. This treats array-like objects like numpy arrays,
+regardless of size.

src/_pytest/python_api.py

@@ -15,9 +15,14 @@
 from typing import Pattern
 from typing import Tuple
 from typing import Type
+from typing import TYPE_CHECKING
 from typing import TypeVar
 from typing import Union
 
+if TYPE_CHECKING:
+    from numpy import ndarray
+
+
 import _pytest._code
 from _pytest.compat import final
 from _pytest.compat import STRING_TYPES
@@ -232,10 +237,11 @@ def __repr__(self) -> str:
     def __eq__(self, actual) -> bool:
         """Return whether the given value is equal to the expected value
         within the pre-specified tolerance."""
-        if _is_numpy_array(actual):
+        asarray = _as_numpy_array(actual)
+        if asarray is not None:
             # Call ``__eq__()`` manually to prevent infinite-recursion with
             # numpy<1.13.  See #3748.
-            return all(self.__eq__(a) for a in actual.flat)
+            return all(self.__eq__(a) for a in asarray.flat)
 
         # Short-circuit exact equality.
         if actual == self.expected:
@@ -521,6 +527,7 @@ def approx(expected, rel=None, abs=None, nan_ok: bool = False) -> ApproxBase:
     elif isinstance(expected, Mapping):
         cls = ApproxMapping
     elif _is_numpy_array(expected):
+        expected = _as_numpy_array(expected)
         cls = ApproxNumpy
     elif (
         isinstance(expected, Iterable)
@@ -536,18 +543,40 @@ def approx(expected, rel=None, abs=None, nan_ok: bool = False) -> ApproxBase:
 
 
 def _is_numpy_array(obj: object) -> bool:
-    """Return true if the given object is a numpy array.
+    """Return true if the given object is implicitly convertible to numpy array.
 
     A special effort is made to avoid importing numpy unless it's really necessary.
     """
     import sys
 
     np: Any = sys.modules.get("numpy")
     if np is not None:
-        return isinstance(obj, np.ndarray)
+        # avoid infinite recursion on numpy scalars, which have __array__
+        if np.isscalar(obj):
+            return False
+        elif isinstance(obj, np.ndarray):
+            return True
+        elif hasattr(obj, "__array__") or hasattr("obj", "__array_interface__"):
+            return True
     return False
 
 
+def _as_numpy_array(obj: object) -> Optional["ndarray"]:
+    """Return an ndarray if obj is implicitly convertible, and numpy is already imported."""
+    import sys
+
+    np: Any = sys.modules.get("numpy")
+    if np is not None:
+        # avoid infinite recursion on numpy scalars, which have __array__
+        if np.isscalar(obj):
+            return None
+        elif isinstance(obj, np.ndarray):
+            return obj
+        elif hasattr(obj, "__array__") or hasattr("obj", "__array_interface__"):
+            return np.asarray(obj)
+    return None
+
+
 # builtin pytest.raises helper
 
 _E = TypeVar("_E", bound=BaseException)

testing/python/approx.py

@@ -447,6 +447,36 @@ def test_numpy_array_wrong_shape(self):
         assert a12 != approx(a21)
         assert a21 != approx(a12)
 
+    def test_numpy_array_protocol(self):
+        """
+        array-like objects such as tensorflow's DeviceArray are handled like ndarray.
+        See issue #8132
+        """
+        np = pytest.importorskip("numpy")
+
+        class DeviceArray:
+            def __init__(self, value, size):
+                self.value = value
+                self.size = size
+
+            def __array__(self):
+                return self.value * np.ones(self.size)
+
+        class DeviceScalar:
+            def __init__(self, value):
+                self.value = value
+
+            def __array__(self):
+                return np.array(self.value)
+
+        expected = 1
+        actual = 1 + 1e-6
+        assert approx(expected) == DeviceArray(actual, size=1)
+        assert approx(expected) == DeviceArray(actual, size=2)
+        assert approx(expected) == DeviceScalar(actual)
+        assert approx(DeviceScalar(expected)) == actual
+        assert approx(DeviceScalar(expected)) == DeviceScalar(actual)
+
     def test_doctests(self, mocked_doctest_runner) -> None:
         import doctest