This documentation is for a development version. Click here for the latest stable release (v1.0.0).


pytest-allclose provides the allclose Pytest fixture, extending numpy.allclose with test-specific features.

A core feature of the allclose fixture is that the tolerances for tests can be configured externally. This allows different repositories to share the same tests, but use different tolerances. See the “Configuration” section below for details.


To use this fixture, install with

pip install pytest-allclose


The allclose fixture is used just like numpy.allclose.

import numpy as np

def test_close(allclose):
    x = np.linspace(-1, 1)
    y = x + 0.001
    assert allclose(y, x, atol=0.002)
    assert not allclose(y, x, atol=0.0005)
    assert not allclose(y, x, rtol=0.002)

Additional arguments

The allclose fixture has a number of arguments that are not part of numpy.allclose. One such argument is xtol, which allows arrays that have been shifted along their first axis by a certain number of steps to be considered close.

import numpy as np

def test_close(allclose):
    x = np.linspace(-1, 1)

    assert allclose(x[1:], x[:-1], xtol=1)
    assert allclose(x[3:], x[:-3], xtol=3)
    assert not allclose(x[3:], x[:-3], xtol=1)

Refer to the allclose API reference for all additional arguments.

RMSE error reporting

The allclose fixture stores root-mean-square error values, which can be reported in the pytest terminal summary. To do so, put the following in your file.

from pytest_allclose import report_rmses

def pytest_terminal_summary(terminalreporter):

See the report_rmses API reference for more information.



allclose_tolerances accepts a list of test name patterns, followed by values for any of the allclose parameters. These values will override any values provided within the test function itself, allowing multiple repositories to use the same test suite, but with different tolerances.

allclose_tolerances = atol=0.3  # set atol for specific test* rtol=0.2  # set rtol for tests matching wildcard* atol=0.1 rtol=0.3  # set both tols for all tests in file
    test_*tion rtol=0.2  # set rtol for all matching tests in any file
    test_function[True] atol=0.1  # set atol only for one parametrization

The only special character recognized in these patterns is the wildcard character *, which matches any group of zero or more characters.

If the test is parametrized, then a pattern like test_name[param0-param1] will match specific parameter settings, and test_name* will match all parameter settings. Note that the latter will match any test that starts with test_name.

If a test has multiple allclose calls, you can use multiple tolerance lines that match the same test to set different values for the first, second, third, etc. calls. If there are more allclose calls than tolerance lines, the last tolerance line will be used for all remaining allclose calls.

Example test file:

def test_close(allclose):
    x = np.linspace(-1, 1)
    y = x + 0.001
    assert allclose(y, x)
    assert not allclose(y, x)

Example configuration file (pytest.ini, setup.cfg):

allclose_tolerances =
    test_close atol=0.002  # affects first allclose call
    test_close atol=0.0005  # affects second allclose call


Different tolerance lines correspond to calls of the function, not lines of code. If you have a for loop that calls allclose 3 times, each of these calls corresponds to a new tolerance line. If you have a fourth allclose call, you would need three tolerance lines for the three calls in the for loop, then a fourth line for the last call.


The patterns for multiple calls of allclose in a function must be exactly the same. This means that if you have specific values for one parametrization and general values for others, you must put the specific values first or they will not have any effect.

Good example, specific takes precedence:

allclose_tolerances =
    test_close[True-1] atol=0.002
    test_close[True-1] atol=0.0005
    test_close* atol=0.001
    test_close* atol=0.0001

Bad example, general takes precedence:

allclose_tolerances =
    test_close* atol=0.001
    test_close* atol=0.0001
    test_close[True-1] atol=0.002
    test_close[True-1] atol=0.0005

API reference


Returns a function checking if two arrays are close, mimicking numpy.allclose.

allclose._allclose(a, b, rtol=1e-5, atol=1e-8, xtol=0, equal_nan=False, print_fail=5, record_rmse=True)

First array to be compared.


Second array to be compared.

rtolfloat, optional

Relative tolerance between a and b (relative to b).

atolfloat, optional

Absolute tolerance between a and b.

xtolint, optional

Allow signals to be right or left shifted by up to xtol indices along the first axis

equal_nanbool, optional

If True, nans will be considered equal to nans.

print_failint, optional

If > 0, print out the first print_fail entries failing the allclose check along the first axis.

record_rmsebool, optional

Whether to record the RMSE value for this comparison. Defaults to True. Set to False whenever a and b should be far apart (when ensuring two signals are sufficiently different, for example).


True if the two arrays are considered close according to the tolerances.

pytest_allclose.report_rmses(terminalreporter, relative=True)[source]

Report RMSEs recorded by the allclose fixture in the Pytest terminal.

This function helps with reporting recorded root mean squared errors (RMSEs). These RMSEs offer a measure of performance for each test by quantifying how close their outputs are to the target values. While this metric has some value on its own, it is most useful as a relative metric, to evaluate if change offers an improvement to tests, and if so, how much.

When using RMSEs, it is important to set record_rmse to False on any allclose call where closer values correspond to a drop in performance (e.g. when using allclose to ensure values are different).


The terminal reporter object provided by pytest_terminal_summary.

relativebool, optional

Whether to print relative (default) or absolute RMSEs. Relative RMSEs are normalized by the mean RMS of a and b in allclose. Since different tests often compare against values of different magnitudes, relative RMSEs provide a better metric across tests by ensuring all tests contribute proportionally to the average RMSE. One exception is when comparing to a signal that is all zeros, since the relative RMSE will always be 2 no matter how close the values are.


See RMSE error reporting for an example.