Software tests

Requirements

The PennyLane test suite requires the Python pytest package, as well as:

  • pytest-cov: determines test coverage

  • pytest-mock: allows replacing components with dummy/mock objects

  • flaky: manages tests with non-deterministic behaviour

These requirements can be installed via pip:

pip install pytest pytest-cov pytest-mock flaky

Creating a test

Every test has to be added to the PennyLane test folder. The test folder follows the structure of the PennyLane module folder. Therefore, tests needs to be added to the corresponding subfolder of the functionality they are testing.

Most tests typically will not require the use of an interface or autodifferentiation framework (such as Autograd, Torch, TensorFlow and Jax). Tests without an interface will be marked as a core test automatically by pytest (the functionality for this is located in conftest.py). For such general tests, you can follow the structure of the example below, where it is recommended that you follow general pytest guidelines:

import pennylane as qml

def test_circuit_expval(self):
    """ Test that the circuit expectation value for PauliX is 0."""
    dev = qml.device("default.qubit", wires=1)

    @qml.qnode(dev)
    def circuit():
        return qml.expval(qml.PauliX(0))

    expval = circuit()

    assert expval == 0

This test will be marked automatically as a core test.

On the other hand, some tests require specific interfaces and need to be marked in order to be run on our Github test suite. Tests involving interfaces have to be marked with their respective marker:

  • @pytest.mark.autograd,

  • @pytest.mark.torch,

  • @pytest.mark.tf, and

  • @pytest.mark.jax.

If tests involve multiple interfaces, one should add the marker:

  • @pytest.mark.all_interfaces.

Warning

Please do not use pytest.importorskip inside your tests. Instead, simply import the autodifferentiation package as needed inside your marked test. The mark will automatically ensure that the test is skipped if the autodifferentiation framework is not installed.

Tests that are not marked but do import an interface will lead to a failure in the GitHub test suite.

Below you can find an example for testing a PennyLane template with Jax:

def circuit_template(features):
    qml.AngleEmbedding(features, range(3))
    return qml.expval(qml.PauliZ(0))

def circuit_decomposed(features):
    qml.RX(features[0], wires=0)
    qml.RX(features[1], wires=1)
    qml.RX(features[2], wires=2)
    return qml.expval(qml.PauliZ(0))

@pytest.mark.jax
def test_jax(self, tol):
    """Tests the jax interface."""

    import jax
    import jax.numpy as jnp

    features = jnp.array([1.0, 1.0, 1.0])

    dev = qml.device("default.qubit", wires=3)

    circuit = qml.QNode(circuit_template, dev, interface="jax")
    circuit2 = qml.QNode(circuit_decomposed, dev, interface="jax")

    res = circuit(features)
    res2 = circuit2(features)
    assert qml.math.allclose(res, res2, atol=tol, rtol=0)

Another example of a test involving multiple interfaces is shown below:

def circuit(features):
    qml.AngleEmbedding(features, range(3))
    return qml.expval(qml.PauliZ(0))

@pytest.mark.all_interfaces
def test_all_interfaces_gradient_agree(self):
    """Test the results are similar between torch and tf"""
    import torch
    import tensorflow as tf

    dev = qml.device("default.qubit", wires=3)

    features_torch = torch.Tensor([1.0, 1.0, 1.0])
    features_tf = tf.Variable([1.0, 1.0, 1.0], dtype=tf.float64)

    circuit_torch = qml.QNode(circuit, dev, interface="torch")
    circuit_tf = qml.QNode(circuit, dev, interface="tf")

    res_torch = circuit_torch(features_torch)
    res_tf = circuit_tf(features_tf)

    assert np.allclose(res_torch, res_tf)

Running the tests

The tests folder of the root PennyLane directory contains the PennyLane test suite. Run all tests in this folder via:

python -m pytest tests

Using python -m ensures that the tests run with the correct Python version if multiple versions are on the system. As the entire test suite takes some time, locally running only relevant files speeds up the debugging cycle. For example, if a developer was adding a new non-parametric operation, they could run:

python -m pytest tests/ops/qubit/test_non_parametric_ops.py

Using pytest -m offers the possibility to select and run tests with specific markers. For example, if Jax is installed and a developer wants to run only Jax related tests, they could run:

python -m pytest tests -m "jax"

There exists markers for interfaces (autograd, torch, tf, jax), for multiple interfaces (all_interfaces) and also for certain PennyLane submodules (qchem and qcut).

For running qchem tests, one can run the following:

python -m pytest tests -m "qchem"

The slowest tests are marked with slow and can be deselected by:

python -m pytest -m "not slow" tests

The pytest -m option supports Boolean combinations of markers. It is therefore possible to run both Jax and TensorFlow tests by writing:

python -m pytest -m "jax and tf" tests

or Jax tests that are not slow:

python -m pytest -m "jax and not slow" tests

Pytest supports many other command-line options, which can be found with the command:

pytest --help

Or by visiting the pytest documentation .

PennyLane provides a set of integration tests for all PennyLane plugins and devices. See the documentation on these tests under the section on the device API. These tests can be run from the PennyLane root folder by:

pytest pennylane/devices/tests --device=default.qubit --shots=1000

All PennyLane tests and the device suite on core devices can be run from the PennyLane root folder via:

make test

Testing Matplotlib based code

Matplotlib images can display differently due to various factors outside the standard developer’s control, such as image backend and available fonts. Even though matplotlib provides functionality for pointwise comparison of images , they require caching correct images in a particular location and are sensitive to details we don’t need to test.

Instead of performing per-pixel comparison of saved images, we can instead inspect the figure and axes objects to ascertain whether they contain the correct information. The figure should contain the axis object in its fig.axes attribute, and the axis object should contain the Artists that get displayed. These artists relevant to us are located in one of three attributes. Each attribute is a list of relevant objects, ordered as they were added:

  • ax.texts

  • ax.lines

  • ax.patches

Instead of testing every relevant piece of information for all objects in the graphic, we can check key pieces of information to make sure everything looks decent. These key pieces of information can include (but are not limited to):

  • number of objects

  • type of objects

  • location

Text objects

Text objects are stored in ax.texts. While the text object has many methods and attributes for relevant information, the two most commonly used in testing text objects are:

  • text_obj.get_text() : Get the string value for the text object

  • text_obj.get_position(): Get the (x,y) position of the object

Lines

2D lines are stored in ax.lines. PennyLane’s circuit drawing code uses lines for wires, SWAP gates, and controlled operations. The most important method for checking lines is line_obj.get_data(). For easier reading, you can also use line_obj.get_xdata() and line_obj.get_ydata().

Patches

Patches can be a wide variety of different objects, like:

Each can have its own getter methods and attributes. For example, an arc has theta1 and theta2. dir(patch_obj) can help developers determine which methods and attributes a given object has.

For Rectangles, the most relevant methods are:

  • rectangle_obj.get_xy()

  • rectangle_obj.get_width()

  • rectangle_obj.get_height()