Contributing to Eradiate#


Eradiate is written and documented in English using Oxford spelling.

Contributing to the documentation#

Building the documentation#

Once Eradiate is installed, the documentation can be built using the following commands:

make docs

After the build is completed, the html document is located in $ERADIATE_SOURCE_DIR/docs/_build/html.


Some parts of the API documentation use static intermediate files generated by a dedicated script. See Building API RST files for more information.

Editing the API documentation#

Our API is documented using docstrings. We follow the Numpy docstring style, with a few changes and updates documented hereafter.


Docstrings start with a newline:

def my_func():
    Docstring contents.

Documenting classes#

In addition to the sections defined in the Numpy style guide, we add a “Fields” section to our class docstrings. Class docstrings therefore have the following structure:

Short summary

A one-line summary that does not use variable names or the function name. It is notably printed in summary tables. See the Numpy docstring style guide for more detail.

Deprecation warning

See the Numpy docstring style guide for more detail.

Extended Summary

See the Numpy docstring style guide for more detail.


Description of the function arguments, keywords and their respective types. This section documents constructor parameters. Note that argument types should reflect types expected by the constructor, which can be broader than field types thanks to the attrs initialisation sequence. See the Numpy docstring style guide for more detail.


Description of class attributes. This section replaces the Attributes section defined in the Numpy style guide. Since Eradiate uses attrs, fields are usually very similar to constructor parameters and rendered with the same style. Types indicated in this section should reflect the true field type, after applying converters. We use dedicated utility functions to generate the Parameters and Fields sections from in-source documentation (see below).

Important don’ts:

  • Properties are documented automatically by the autosummary extension: do not document them in this section, they will be displayed in a dedicated Attributes rubric on the class documentation page.

  • Do not use ivar to document attributes: use this section instead.

  • Do not use the Methods section.


See the Numpy docstring style guide for more detail.


See the Numpy docstring style guide for more detail.


See the Numpy docstring style guide for more detail.

Other Parameters

See the Numpy docstring style guide for more detail.


See the Numpy docstring style guide for more detail.


See the Numpy docstring style guide for more detail.


See the Numpy docstring style guide for more detail.

See Also

See the Numpy docstring style guide for more detail.


See the Numpy docstring style guide for more detail.


See the Numpy docstring style guide for more detail.


See the Numpy docstring style guide for more detail.


See the Numpy docstring style guide for more detail.

Field documentation helpers#

Fields are documented using specific helper functions provided as part of Eradiate’ documentation framework. They notably allow to automatically create class docstrings for classes with inherited fields.

The parse_docs() decorator must be applied to the documented class prior to any other action. Then, each declared attribute can be documented using the documented() function:

import attr
from typing import Optional
from eradiate.util.attrs import parse_docs, documented

@parse_docs  # Must be applied **after** attr.s
class MyClass:
    field: Optional[float] = documented(
        doc="A documented attribute",
        type="float, optional",

In addition, a init_type argument lets the user specify if constructor argument types are different from the field type. This is particularly useful when a converter is systematically applied to field values upon initialisation:

import attr
import numpy as np
from eradiate.util.attrs import parse_docs, documented

@parse_docs  # Must be applied **after** attr.s
class MyClass:
    field: np.ndarray = documented(
        doc="A documented attribute",

The doc, type, init_type and default parameters currently only support string values.

Fields are sometimes partially redefined, but parts of their documentation can be reused. For such cases, we provide the get_doc() function:

import attr
from eradiate.util.attrs parse_docs, documented, import get_doc

class MyChildClass(MyClass):
    field = documented(
        doc=get_doc(MyClass, "field", "doc"),
        type=get_doc(MyClass, "field", "type"),

Building API RST files#

Parts of the API documentation are generated using a dedicated Python script. The generation process is integrated in the Sphinx configuration, but it can sometimes be useful to build those static files manually. This can be done with the docs-rst make target:

make docs-rst

Editing tutorials#

Eradiate comes with tutorials shipped as Jupyter notebooks, saved to the “tutorials“ submodule. They are integrated in this documentation using the nbsphinx extension.

The recommended way to edit tutorials is as follows:

  1. Open a terminal and start a Jupyter session.

  2. In another terminal, open a Sphinx server using the following command at the root of your local copy of Eradiate:

    make docs-serve
  3. Browse to the tutorial you want to edit or create a new one using the tutorial template. You can now edit the content and see how it renders dynamically.


    Make sure that the first cell is as follows:

    %reload_ext eradiate.notebook.tutorials

Nbsphinx renders markdown cells, but also allows to define raw reST cells, which then support all usual Sphinx features (references, admonitions, etc.). See the documentation for more detail.

Tutorials are currently not run as part of the documentation build process; instead, the output of the rendered notebook is checked in to the Git repository. The reason for this is that rendering tutorials when building the documentation would require a fully functional copy of Eradiate, including its radiometric kernel Mitsuba. This is currently unachievable on the Read the Docs service we use to deploy automatically the documentation upon committing to GitHub: Mitsuba must be compiled and Read the Docs does not support its build process.


Once you’re done editing a tutorial, do not forget to rerun it entirely in a clean Jupyter session to render it as if you were a user.

Thumbnail galleries are not trivial difficult to fine-tune. The following pages are useful when working on them:

Contributing to the code base#


  • The Eradiate codebase is written following Python’s PEP8. Its code formatter of choice is Black and its linter of choice is ruff, for which a configuration is provided as part of the pyproject.toml file. Editor integration instructions are available for Black and for ruff. Both tools are part of our pre-commit hook set, which we strong recommend to install.

  • We write our docstrings following the Numpydoc format. We use the """-on-separate-lines style:

    def func(x):
        Do something.
        Further detail on what this function does.
  • We use type hints in our library code. We do not use type hints in test code in general.

Code writing#


  • Eradiate is built using the attrs library. It is strongly recommended to read the attrs documentation prior to writing additional classes. In particular, it is important to understand the attrs initialisation sequence, as well as how callables can be used to set defaults and to create converters and validators.

  • Eradiate’s unit handling is based on Pint, whose documentation is also a very helpful read.

  • Eradiate uses custom Pint-based extensions to attrs now developed as the standalone project Pinttrs. Reading the Pinttrs docs is highly recommended.

  • Eradiate uses factories based on the Dessine-moi library. Reading the Dessine-moi docs is recommended.

When writing code for Eradiate, the following conventions and practices should be followed.

Prefer relative imports in library code

We generally use relative imports in library code, and absolute imports in tests and application code.

Minimise class initialisation code

Using attrs for class writing encourages to minimise the amount of complex logic implemented by constructors. Although attrs provides the __attrs_post_init__() method to do so, we try to avoid it as much as possible. If a constructor must perform special tasks, then this logic is usually better implemented as a class method constructor (e.g. from_something()).

Initialisation from dictionaries

A lot of Eradiate’s classes can be instantiated using dictionaries. Most of them leverage factories for that purpose (see Factory guide). This, in practice, reserves the "type" and "construct" parameters, meaning that factory-registered classes cannot have type or construct fields.

For classes unregistered to any factory, our convention is to implement dictionary-based initialisation as a from_dict() class method constructor. It should implement behaviour similar to what Factory.convert() does, i.e.:

  • interpret units using pinttr.interpret_units();

  • [optional] if relevant, allow for class method constructor selection using the "construct" parameter.

Deprecations and removals#

Eradiate tries to remain backward-compatible when possible. Sometimes however, compatibility must be broken. Following the recommended practice in the Python community, removals are, whenever possible, preceded by a deprecation period during which a deprecated component is still available, marked as such in the documentation, and using it triggers a DeprecationWarning.

This workflow is facilitated by components defined in the util.deprecation module, and in particular the deprecated() decorator. Be sure to use them when relevant.


Eradiate is shipped with a series of tests written with pytest.

At the highest level, there is a separation of tests for Mitsuba plugins which are maintained in the Eradiate codebase and tests for Eradiate’s high-level code. The tests for Eradiate are then grouped by complexity. First unit tests are executed, followed by system tests and finally regression tests.

Running the test suite#

To run the test suite, invoke pytest with the following command:

pytest tests

Testing guidelines#

Writing test specification#

Eradiate’s tests can be roughly categorised as follows:

  • unit tests focus on the smallest testable units of code;

  • system tests check the behaviour of entire applications;

  • regression tests which compare simulation results with previous versions.

While categorising each individual test is not always an easy task, this nomenclature highlights the fact that tests have varied degrees of complexity. When the rationale, setup and course of action of a test is not obvious by reading the corresponding source code, properly documenting it in a structured way is crucial. For this reason, Eradiate defines a test description template to be used for system and regression tests.

The test specification consists of three main parts:

  1. the description of the test rationale;

  2. the details of the setup, explaining, in prose, how a test is designed;

  3. the expected outcome of the test, which describes based on what the test should pass or fail.

The following template can be copied to new test cases and the information filled in as needed. Note that we strongly suggest using string literals (prefixed with a r) in order to avoid issues with escape sequences.

Test title

:Description: This is the short description of the test case


This is some explanatory text

* This section explains the details
* Of how the test is implemented
* It can contain math! :math:`e^{i\pi}=-1`

Expected behaviour

This section explains the expected result of the test and how it is asserted.

* We assert that something was calculated
* Additionally the result must be correct

The test specification can hold any valid restructured text. A quick rundown on that can be found here .

Regression tests#

Eradiate’s regression tests are designed to allow the monitoring of results over time. Each test produces a NetCDF file with the current results as well as an image containing plots and metrics, comparing the current version of Eradiate to the reference results. The results of these tests can be archived for future reference.

To run the regression tests isolated from the rest of the test suite, we introduced the regression fixture. To run only the regression tests, invoke pytest like this:

pytest tests -m "regression" --artefact-dir <a directory of your choice>

The artefact_dir parameter defines the output directory in which the results and plots will be placed. If the directory does not exist, it will be created. The artefact directory defaults to ./build/test_artefacts, which is resolved relative to the current working directory.

Adding new regression tests#

Regression tests use a comparison framework providing interfaces for statistical and other metric-based tests. Relevant components are listed in the API reference [eradiate.test_tools].

These tests are based on comparing the results of a computation to a reference, computed on a previous version of the code which was deemed correct by other means.

To implement tests based on this framework, we provide helper classes which can be imported from the eradiate.test_tools.regression module:

import eradiate.test_tools.regression as ttr

Within your test case, you then instantiate one of the subclasses:

result = your_eradiate_simulation()

test = ttr.Chi2Test(

After running a simulation on an Eradiate scene, you provide the resulting dataset as well as a path to the reference result to the helper class. Adding a threshold value, which may depend on the scenario and the chosen metric, and a path and filename for the outputs generated by the class the test is ready. To execute the test it exposes the method, which handles computing the metric, storing the results in the given path, and returns the test outcome as a boolean.

The test will store two NetCDF files and an image file with a visualisation of the results in the directory given as archive_filename. It will store the new result and the reference in two files, adding -result and -ref suffixes to the provided filename.

To handle the test result simply use an assertion:


Analysing the results#

If the test fails due to a significant difference between the reference and the result the output can help in analysis. The reference data and the result are stored in two NetCDF files under the path given in archive_filename, which can be imported and used in python scripts for detailed analysis. Furthermore the test adds an overview plot made up of four parts: A direct visualisation of the result and reference data on the same axis, the absolute and relative difference between result and reference in their own axes and the numerical value of the chosen metric.

In case this difference stems from a change made to Eradiate, which significantly alters the code’s behaviour, the reference needs to be updated. In this case, replace the existing reference file in the data repository and create a pull request for the maintainers to review and add.

In case the test fails due to a missing or non found reference, for example when adding a new test case, the helper will not attempt to compute the metric at all. Instead it will output the simulation result as NetCDF under the given path with the -ref suffix alongside a simple visualisation of the result. The output can then be added to the data repository as mentioned above.

Test report#

Optionally, test results may be visualised using a report generated with a tool located on a dedicated repository.

The report summarises test outcomes and generates detailed entries for tests specified with the docstring format specified above.

The test specification of unit tests is not parsed for the test report and does not have to comply with these guidelines. For those, a short explanation is sufficient, but the three general parts mentioned above should still serve as a guideline for relevant and helpful test specification.


Shallow submodule caveats#

Eradiate uses Git submodules to ship some of its data. Over time, these can grow and become large enough so that using a shallow submodule. Shallow clones do not contain the entire history of the repository and are therefore more lightweight, saving bandwidth upon cloning.

However, shallow clones can be difficult to work with, especially when one starts branching. If a shallow submodule is missing a remote branch you’d expect it to track, this post contains probably what you need to do:

cd my-shallow-submodule
git remote set-branches origin '*'
git fetch -v
git checkout the-branch-i-ve-been-looking-for


Tests are a very opportunity to profile Eradiate. We recommend running tests with pytest-profiling (see documentation for usage instructions, it’s basically about installing the package then running pytest with the --profile option).

Profiling stats can then be visualised with SnakeViz.