Running and writing tests

Writing tests

Every new function that is written in to HyperSpy should be tested and documented. HyperSpy uses the pytest library for testing. The tests reside in the hyperspy.tests module.

Tests are short functions, found in hyperspy/tests, that call your functions under some known conditions and check the outputs against known values. They should depend on as few other features as possible so that when they break, we know exactly what caused it. Ideally, the tests should be written at the same time as the code itself, as they are very convenient to run to check outputs when coding. Writing tests can seem laborious but you’ll probably soon find that they are very important, as they force you to sanity-check all the work that you do.

Useful testing hints:

  • When comparing integers, it’s fine to use ==

  • When comparing floats, be sure to use np.testing.assert_allclose() or np.testing.assert_almost_equal()

  • np.testing.assert_allclose() is also convenient for comparing numpy arrays

  • The hyperspy.misc.test_utils.py contains a few useful functions for testing

  • @pytest.mark.parametrize() is a very convenient decorator to test several parameters of the same function without having to write to much repetitive code, which is often error-prone. See pytest documentation for more details.

  • It is good to check that the tests do not use too much memory after creating new tests. If you need to explicitly delete your objects and free memory, you can do the following to release the memory associated with the s object:

>>> import gc

>>> # Do some work with the object
>>> s = ...

>>> # Clear the memory
>>> del s
>>> gc.collect()

Running tests

First ensure pytest and its plugins are installed by:

# If using a standard hyperspy install
$ pip install hyperspy[tests]

# Or, from a hyperspy local development directory
$ pip install -e .[tests]

# Or just installing the dependencies using conda
$ conda install -c conda-forge pytest pytest-mpl

To run them:

$ pytest --mpl --pyargs hyperspy

Or, from HyperSpy’s project folder, simply:

$ pytest

Note

pytest configuration options are set in the setup.cfg file, under the [tool:pytest] section. See the pytest configuration documentation for more details.

The HyperSpy test suite can also be run in parallel if you have multiple CPUs available, using the pytest-xdist plugin. If you have the plugin installed, HyperSpy will automatically run the test suite in parallel on your machine.

# To run on all the cores of your machine
$ pytest -n auto --dist loadfile

# To run on 2 cores
$ pytest -n 2 --dist loadfile

The --dist loadfile argument will group tests by their containing file. The groups are then distributed to available workers as whole units, thus guaranteeing that all tests in a file run in the same worker.

Note

Running tests in parallel using pytest-xdist will change the content and format of the output of pytest to the console. We recommend installing pytest-sugar to produce nicer-looking output including an animated progressbar.

Flaky tests

Test functions can sometimes exhibit intermittent or sporadic failure, with seemingly random or non-deterministic behaviour. They may sometimes pass or sometimes fail, and it won’t always be clear why. These are usually known as “flaky” tests.

One way to approach flaky tests is to rerun them, to see if the failure was a one-off. This can be achieved using the pytest-rerunfailures plugin.

# To re-run all test suite failures a maximum of 3 times
$ pytest --reruns 3

# To wait 1 second before the next retry
$ pytest --reruns 3 --reruns-delay 1

You can read more about flaky tests in the pytest documentation.

Test coverage

Once you have pushed your pull request to the official HyperSpy repository, you can see the coverage of your tests using the codecov.io check for your PR. There should be a link to it at the bottom of your PR on the Github PR page. This service can help you to find how well your code is being tested and exactly which parts are not currently tested.

You can also measure code coverage locally. If you have installed pytest-cov, you can run (from HyperSpy’s project folder):

$ pytest --cov=hyperspy

Configuration options for code coverage are also set in the setup.cfg file, under the [coverage:run] and [coverage:report] sections. See the coverage documentation for more details.

Note

The codecov.io check in your PR will fail if it either decreases the overall test coverage of HyperSpy, or if any of the lines introduced in your diff are not covered.

Continuous integration (CI)

The HyperSpy test suite is run using continuous integration services provided by Github Actions and Azure Pipelines. In case of Azure Pipelines, CI helper scripts are pulled from the ci-scripts repository.

The testing matrix is as follows:

  • Github Actions: test a range of Python versions on Linux, MacOS and Windows; all dependencies are installed from PyPI. See .github/workflows/tests.yml in the HyperSpy repository for further details.

  • Azure Pipeline: test a range of Python versions on Linux, MacOS and Windows; all dependencies are installed from Anaconda Cloud using the “conda-forge” channel. See azure-pipelines.yml in the HyperSpy repository for further details.

  • The testing of HyperSpy extensions is described in the integration test suite section.

This testing matrix has been designed to be simple and easy to maintain, whilst ensuring that packages from PyPI and Anaconda cloud are not mixed in order to avoid red herring failures of the test suite caused by application binary interface (ABI) incompatibility between dependencies.

The most recent versions of packages are usually available first on PyPI, before they are available on Anaconda Cloud. These means that if a recent release of a dependency breaks the test suite, it should happen first on Github Actions. Similarly, deprecation warnings will usually appear first on Github Actions.

The documentation build is done on both Github Actions and Read the Docs, and it is worth checking that no new warnings have been introduced when writing documentation in the user guide or in the docstrings.

The Github Actions testing matrix also includes the following special cases:

  • The test suite is run against HyperSpy’s minimum requirements on Python 3.7 on Linux. This will skip any tests that require optional packages such as scikit-learn.

  • The test suite is run against the oldest supported versions of numpy, matplotlib and scipy. For more details, see this Github issue.

  • The test suite is run against the development supported versions of numpy, scipy, scikit-learn and scikit-image using the weekly build wheels available on https://anaconda.org/scipy-wheels-nightly. For more details, see this Github issue.

Plot testing

Plotting is tested using the @pytest.mark.mpl_image_compare decorator of the pytest mpl plugin. This decorator uses reference images to compare with the generated output during the tests. The reference images are located in the folder defined by the argument baseline_dir of the @pytest.mark.mpl_image_compare decorator.

To run plot tests, you simply need to add the option --mpl:

$ pytest --mpl

If you don’t use --mpl, the test functions will be executed, but the images will not be compared to the reference images.

If you need to add or change some plots, follow the workflow below:

  1. Write the tests using appropriate decorators such as @pytest.mark.mpl_image_compare.

  2. If you need to generate a new reference image in the folder plot_test_dir, for example, run: pytest --mpl-generate-path=plot_test_dir

  3. Run again the tests and this time they should pass.

  4. Use git add to put the new file in the git repository.

When the plotting tests fail, it is possible to download the figure comparison images generated by pytest-mpl in the artifacts tabs of the corresponding build on Azure Pipeliness:

../_images/azure_pipeline_artifacts.png

The plotting tests are tested on Azure Pipelines against a specific version of matplotlib defined in conda_environment_dev.yml. This is because small changes in the way matplotlib generates the figure between versions can sometimes make the tests fail.

For plotting tests, the matplotlib backend is set to agg by setting the MPLBACKEND environment variable to agg. At the first import of matplotlib.pyplot, matplotlib will look at the MPLBACKEND environment variable and accordingly set the backend.

Exporting pytest results as HTML

With pytest-html, it is possible to export the results of running pytest for easier viewing. It can be installed by conda:

$ conda install pytest-html

and run by:

$ pytest --mpl --html=report.html

See pytest-mpl for more details.