Attribute error in pytest.approx for types implicitly convertible to numpy arrays

[ascended121]

• a detailed description of the bug or problem you are having

When using pytest.approx with custom types which are implicitly convertible to numpy arrays, the following error is observed:

E       AssertionError: assert <test_matrix_...x7f9169ac4ca0> == approx([1.0 ±....0 ± 4.0e-06])
E         (pytest_assertion plugin: representation of details failed: ~/.venv/lib/python3.10/site-packages/_pytest/python_api.py:175: AttributeError: 'MyType' object has no attribute 'shape'.
E          Probably an object has a faulty __repr__.)

Note: The line number in the error message is a little off from the current head. That line is here.

Looking at the code, pytest.approx sensibly chooses the numpy implementation ApproxNumpy, because the expected value is implicitly convertible to a numpy array. However, the shape check assumes that other_side is a numpy array (rather than assuming its implicitly convertible to a numpy array, like the expected value is) and so assumes that it has a shape attribute (which is not part of how approx determines if something is a numpy array).

It seems like either:

1. ApproxNumpy._repr_compare() needs to convert other_side to a numpy array (ie _as_numpy_array(other_side))
2. _as_numpy_array needs to check if the object has a shape attribute, if the code is going to assume that's the case

• output of pip list from the virtual environment you are using:
  Relevant packages:
  numpy: 1.26.1

• pytest and operating system versions

pytest: 7.4.3
Ubuntu: 22.04

• minimal example if possible

import numpy as np
import pytest

class MyType:
    """Type which is implicitly convertible to a numpy array."""
    def __init__(self, vals):
        self.vals = vals

    def __repr__(self):
        return f"{self.__class__.__name__}({self.vals})"

    def __array__(self, dtype=None, copy=None):
        return np.array(self.vals)

def test_approx_eq():
    vec1 = MyType([1.0, 2.0, 3.0])
    vec2 = MyType([1.0, 2.0, 4.0])
    assert vec1 == pytest.approx(vec2)

[poulami-sau]

I can try working on this!

[poulami-sau]

I'm not able to reproduce this error in the test cases for testing/python/approx.py, which is what I need to do to verify it in tox. Do you have any pointers?

This is one way I wrote the test case:

class ImplicitArray:
    """Type which is implicitly convertible to a numpy array."""
    def __init__(self, vals):
        self.vals = vals

    def __repr__(self):
        return f"{self.__class__.__name__}({self.vals})"

    def __array__(self, dtype=None, copy=None):
        print("called __array__ in ImplicitArray")
        np = pytest.importorskip("numpy")
        return np.array(self.vals)

class TestApprox:
    def test_numpy_array_implicit_conversion(self):
        vec1 = ImplicitArray([1.0, 2.0, 3.0])
        vec2 = ImplicitArray([1.0, 2.0, 4.0])
        assert vec1 == approx(vec2)

Here's another way that I wrote it:

class TestApprox:
    def test_numpy_array_implicit_conversion(self):
        np = pytest.importorskip("numpy")

        class ImplicitArray:
            """Type which is implicitly convertible to a numpy array."""
            def __init__(self, vals):
                self.vals = vals

            def __repr__(self):
                return f"{self.__class__.__name__}({self.vals})"

            def __array__(self, dtype=None, copy=None):
                print("called __array__ in ImplicitArray")
                return np.array(self.vals)

        vec1 = ImplicitArray([1.0, 2.0, 3.0])
        vec2 = ImplicitArray([1.0, 2.0, 4.0])
        assert vec1 == approx(vec2)

But for both cases, they are passing when I know it shouldn't, and I haven't made any edits to ApproxNumpy yet.

[poulami-sau]

Okay I think tox doesn't work for this as I cannot verify if numpy was actually installed or not with tox. However, it works with the alternate method for testing and I've fixed the bug. I can submit a pull request soon!

[nicoddemus]

Fixed in #12232