Practices#
Purpose: In the development of SimPEG, we strive to follow best practices. Here, we provide an overview of those practices and some tools we use to support them.
Here we cover
Testing#
On each update, SimPEG is tested using the continuous integration service Azure pipelines. We use Codecov to check and provide stats on how much of the code base is covered by tests. This tells which lines of code have been run in the test suite. It does not tell you about the quality of the tests run! In order to assess that, have a look at the tests we are running - they tell you the assumptions that we do not want to break within the code base.
Within the repository, the tests are located in the top-level tests directory. Tests are organized similar to the structure of the repository. There are several types of tests we employ, this is not an exhaustive list, but meant to provide a few places to look when you are developing and would like to check that the code you wrote satisfies the assumptions you think it should.
Testing is performed with pytest which is available through PyPI.
Checkout the docs on pytest.
Compare with known values#
In a simple case, you might know the exact value of what the output should be
and you can assert that this is in fact the case. For example,
we setup a 3D BaseRectangularMesh and assert that it has 3 dimensions.
from discretize.base import BaseRectangularMesh
import numpy as np
mesh = BaseRectangularMesh([6, 2, 3])
def test_mesh_dimensions():
assert mesh.dim == 3
All functions with the naming convention test_XXX
are run. Here we check that the dimensions are correct for the 3D mesh.
If the value is not an integer, you can be subject to floating point errors,
so assert == might be too harsh. In this case, you will want to use
the numpy.testing module to check for approximate equals. For instance,
import numpy as np
import discretize
from SimPEG import maps
def test_map_multiplication(self):
mesh = discretize.TensorMesh([2,3])
exp_map = maps.ExpMap(mesh)
vert_map = maps.SurjectVertical1D(mesh)
combo = exp_map*vert_map
m = np.arange(3.0)
t_true = np.exp(np.r_[0,0,1,1,2,2.])
np.testing.assert_allclose(combo * m, t_true)
These are rather simple examples, more advanced tests might include solving an
electromagnetic problem numerically and comparing it to an analytical
solution , or
performing an adjoint test to test Jvec and Jtvec.
Order and Derivative Tests#
Order tests can be used when you are testing differential operators (we are using a second-order, staggered grid discretization for our operators). For example, testing a 2D curl operator in test_operators.py
import numpy as np
import unittest
from discretize.tests import OrderTest
class TestCurl2D(OrderTest):
name = "Cell Grad 2D - Dirichlet"
meshTypes = ['uniformTensorMesh']
meshDimension = 2
meshSizes = [8, 16, 32, 64]
def getError(self):
# Test function
ex = lambda x, y: np.cos(y)
ey = lambda x, y: np.cos(x)
sol = lambda x, y: -np.sin(x)+np.sin(y)
sol_curl2d = call2(sol, self.M.gridCC)
Ec = cartE2(self.M, ex, ey)
sol_ana = self.M.edge_curl*self.M.project_face_vector(Ec)
err = np.linalg.norm((sol_curl2d-sol_ana), np.inf)
return err
def test_order(self):
self.orderTest()
Derivative tests are a particular type or Order and Derivative Tests, and since they
are used so extensively, SimPEG includes a check_derivative method.
In the case of testing a derivative, we consider a Taylor expansion of a function about \(x\). For a small perturbation \(\Delta x\),
As \(\Delta x\) decreases, we expect \(\|f(x) - f(x + \Delta x)\|\) to have first order convergence (e.g. the improvement in the approximation is directly related to how small \(\Delta x\) is, while if we include the first derivative in our approximation, we expect that \(\|f(x) + J(x)\Delta x - f(x + \Delta x)\|\) to converge at a second-order rate. For example, all maps have an associated derivative test . An example from test_FDEM_derivs.py
def deriv_test(fdemType, comp):
# setup simulation, survey
def fun(x):
return survey.dpred(x), lambda x: sim.Jvec(x0, x)
return tests.check_derivative(fun, x0, num=2, plotIt=False, eps=FLR)
Documentation#
Documentation helps others use your code! Please document new contributions. SimPEG tries to follow the numpydoc style of docstrings (check out the style guide). SimPEG then uses sphinx to build the documentation. When documenting a new class or function, please include a description (with math if it solves an equation), inputs, outputs and preferably a small example.
For example:
class WeightedLeastSquares(BaseComboRegularization):
r"""Weighted least squares measure on model smallness and smoothness.
L2 regularization with both smallness and smoothness (first order
derivative) contributions.
Parameters
----------
mesh : discretize.base.BaseMesh
The mesh on which the model parameters are defined. This is used
for constructing difference operators for the smoothness terms.
active_cells : array_like of bool or int, optional
List of active cell indices, or a `mesh.n_cells` boolean array
describing active cells.
alpha_s : float, optional
Smallness weight
alpha_x, alpha_y, alpha_z : float or None, optional
First order smoothness weights for the respective dimensions.
`None` implies setting these weights using the `length_scale`
parameters.
alpha_xx, alpha_yy, alpha_zz : float, optional
Second order smoothness weights for the respective dimensions.
length_scale_x, length_scale_y, length_scale_z : float, optional
First order smoothness length scales for the respective dimensions.
mapping : SimPEG.maps.IdentityMap, optional
A mapping to apply to the model before regularization.
reference_model : array_like, optional
reference_model_in_smooth : bool, optional
Whether to include the reference model in the smoothness terms.
weights : None, array_like, or dict or array_like, optional
User defined weights. It is recommended to interact with weights using
the `get_weights`, `set_weights` functionality.
Notes
-----
The function defined here approximates:
.. math::
\phi_m(\mathbf{m}) = \alpha_s \| W_s (\mathbf{m} - \mathbf{m_{ref}} ) \|^2
+ \alpha_x \| W_x \frac{\partial}{\partial x} (\mathbf{m} - \mathbf{m_{ref}} ) \|^2
+ \alpha_y \| W_y \frac{\partial}{\partial y} (\mathbf{m} - \mathbf{m_{ref}} ) \|^2
+ \alpha_z \| W_z \frac{\partial}{\partial z} (\mathbf{m} - \mathbf{m_{ref}} ) \|^2
Note if the key word argument `reference_model_in_smooth` is False, then mref is not
included in the smoothness contribution.
If length scales are used to set the smoothness weights, alphas are respectively set internally using:
>>> alpha_x = (length_scale_x * min(mesh.edge_lengths)) ** 2
"""
Style#
Consistency makes code more readable and easier for collaborators to jump in. PEP 8 provides conventions for coding in Python. SimPEG is currently not PEP 8 compliant, but we are working towards it and would appreciate contributions that do too! Often, most python text editors can be configured to issue warnings for non-compliant styles.
SimPEG uses black version 23.1.0 to autoformat the code base, and all additions to the code are tested to ensure that they are compliant with black. We recommend installing pre-commit hooks that are run on every commit to automatically ensure compliance.
We also actively update the code base to ensure pep8 compliance by checking with flake8 This performs style checks that could lead towards bugs, performs checks on consistent documentation formatting, or just identify poor coding practices. This is an ongoing process where we are fixing one style warning at a time. The fixed style warnings are checked to ensure no new code goes against an already established style. This test can also be installed locally using pre-commit hooks, similar to black above.
Licensing#
We want SimPEG to be a useful resource for the geoscience community and believe that following open development practices is the best way to do that. SimPEG is licensed under the MIT license which is allows open and commercial use and extension of SimPEG. It does not force packages that use SimPEG to be open source nor does it restrict commercial use.