If you have issues when using Nilearn, or if you have questions on how to use it, please don't hesitate to reach out!
There are currently three ways to interact with the Nilearn team: through the :neurostars:`neurostars <>` forum, our :nilearn-gh:`github <>` issues, and through our weekly :nilearn-gh:`office hours <issues/2882>`, usually every Friday from 4pm to 5pm UTC.
If you have a usage question, that is if you need help troubleshooting scripts using Nilearn, we would appreciate it if you either ask it during office hours or create a topic on :neurostars:`neurostars <>` with the "nilearn" tag. Asking questions or reporting issues is always valuable because it will help other users having the same problem. So, please don't hold onto a burning question!
We ask that you don't open an issue on :nilearn-gh:`GitHub <>` for usage questions. We use our :nilearn-gh:`GitHub <>` issue board for bug reports, feature requests, and documentation changes only.
If you are interested in contributing to the Nilearn project, we thank you very much. Note that there are multiple ways to help us, and not all of them require writing code.
We welcome open discussion around improvements---both to the documentation as well as to the code base---through our GitHub issue board!
- If you think you have discovered a bug, please start by searching through the existing :nilearn-gh:`issues <issues>` to make sure it has not already been reported. If the bug has not been reported yet, create an :nilearn-gh:`new issue <issues/new/choose>` including a minimal runnable example to showcase it (using Nilearn data) as well as your OS and Nilearn version.
- If you have an idea for a new feature, check if it is in the :ref:`nilearn_scope` and feel free to open a :nilearn-gh:`new issue <issues/new/choose>` to discuss it.
- If you think the documentation can be improved, please open a :nilearn-gh:`new issue <issues/new/choose>` to discuss what you would like to change! This helps to confirm that your proposed improvements don't overlap with any ongoing work.
Another way to help the project is to answer questions on :neurostars:`neurostars <>`, or comment on github :nilearn-gh:`issues <issues>`. Some :nilearn-gh:`issues <issues>` are used to gather user opinions on various questions, and any input from the community is valuable to us.
Any addition to the Nilearn's code base has to be reviewed and approved by several people including at least two :ref:`core_devs`. This can put a heavy burden on :ref:`core_devs` when a lot of :nilearn-gh:`pull requests <pulls>` are opened at the same time. We welcome help in reviewing :nilearn-gh:`pull requests <pulls>` from any community member. We do not expect community members to be experts in all changes included in :nilearn-gh:`pull requests <pulls>`, and we encourage you to concentrate on those code changes that you feel comfortable with. As always, more eyes on a code change means that the code is more likely to work in a wide variety of contexts!
The :ref:`triage` is composed of community members who have permission on :nilearn-gh:`github <>` to label and close :nilearn-gh:`issues <issues>`. Their work is crucial to improve the communication in the project and limit the crowding of the issue tracker. The :ref:`issue_labels` and :ref:`closing_policy` of the project is defined in more details in the :ref:`maintenance_process` page.
Any community member may become a member of the Nilearn :ref:`triage`. If you are interested in joining the :ref:`triage`, please read the :ref:`maintenance_process` page and contact one of the :ref:`core_devs` to express your interest!
If you want to contribute code:
- For new features, please be sure to create a :nilearn-gh:`new issue <issues/new/choose>` first, to discuss whether it can be included and its specifications.
- To help with known :nilearn-gh:`issues <issues>`, please check :nilearn-gh:`good first issues <labels/Good%20first%20issue>` to get started, :nilearn-gh:`known bugs <labels/Bug>`, or :nilearn-gh:`proposed enhancements <labels/Enhancement>`.
Please see the :ref:`contributing_code` section for more detailed information, including instructions for Setting up your environment and a description of the Contribution Guidelines.
The following sections explain the :ref:`nilearn_scope` and :ref:`nilearn_governance`, which jointly determine whether potential contributions will be accepted into the project.
Nilearn is an Open-source Python package for visualizing and analyzing human brain MRI data. It provides statistical and machine-learning tools for brain mapping, connectivity estimation and predictive modelling. It brings visualization tools with instructive documentation & open community.
Nilearn targets ease of use, but as Python code. In other words, we will not add graphical user interfaces, but we want our code to be as easy to understand as possible, with easy prototyping and debugging, even for beginners in Python.
We are parsimonious in the way we add features to the project, as it puts on weight. To assess new features, our main concern is their usefulness to a number of our users. To make Nilearn high-quality and sustainable we also weigh their benefits (i.e., new features, ease of use) with their cost (i.e., complexity of the code, runtime of the examples). As a rule of thumb:
- To be accepted, new features must be in the scope of the project and correspond to an established practice (typically as used in scientific publications).
- It must have a concrete use case, illustrated with a simple example in the Nilearn documentation to teach it easily to end-users.
- It must be thoroughly tested, and respect coding conventions of the existing codebase.
- Features introducing new dependencies will generally not be accepted.
- Downloaders for new atlases are welcome if they comes with an example.
- Downloaders for new datasets are usually discouraged. We will consider adding fetchers only for light datasets which are needed to demo and teach features.
Exhaustive criteria used in the review process are detailed in the contribution guide below. Be sure to read and follow them so that your code can be accepted quickly.
We strongly aim to be a community oriented project where decisions are made based on consensus according to the criteria described above. Discussions are public, held on :nilearn-gh:`issues <issues>` and :nilearn-gh:`pull requests <pulls>` in Github. All modifications of the codebase are ultimately checked during a reviewing process, where maintainers or contributors make sure they respect the :ref:`contribution_guidelines`. To be merged, a pull request usually needs to be accepted by two maintainers. In case a consensus does not emerge easily, the decisions are made by the :ref:`core_devs`, i.e., people with write access to the repository, as listed :ref:`here <core_devs>`.
This project, hosted on :nilearn-gh:`\ `, is a community effort, and everyone is welcome to contribute. We value very much your feedback and opinion on features that should be improved or added. All discussions are public and held on relevant :nilearn-gh:`issues <issues>` or :nilearn-gh:`pull requests <pulls>`. To discuss your matter, please comment on a relevant :nilearn-gh:`issue <issues>` or open a new one.
The best way to contribute and to help the project is to start working on known :nilearn-gh:`issues <issues>` such as :nilearn-gh:`good first issues <labels/Good%20first%20issue>`, :nilearn-gh:`known bugs <labels/Bug>` or :nilearn-gh:`proposed enhancements <labels/Enhancement>`. If an issue does not already exist for a potential contribution, we ask that you first open a :nilearn-gh:`new issue <issues/new/choose>` before sending a :ref:`pull request` to discuss scope and potential design choices in advance.
When modifying the codebase, we ask every contributor to respect common guidelines. Those are inspired from :sklearn:`scikit-learn <developers/contributing.html#contributing-code>` and ensure Nilearn remains simple to understand, efficient and maintainable. For example, code needs to be tested and those tests need to run quickly in order not to burden the development process. To keep continuous integration efficient with our limited infrastructure, running all the examples must lead to downloading a limited amount of data (gigabytes) and execute in a reasonable amount of time (less than an hour). Those guidelines will hence be enforced during the reviewing process. The section Setting up your environment will help you to quickly get familiar with the tools we use for development and deployment.
| Which PR ? | Guidelines | |
|---|---|---|
| PR Structure | Any |
|
| Coding Style | Any |
|
| Tests |
|
|
| Examples | New features |
|
| Documentation | Any |
|
A new pull request must have a clear scope, conveyed through its name, a reference to the issue it targets (through the exact mention "Closes #XXXX"), and a synthetic summary of its goals and main steps. When working on big contributions, we advise contributors to split them into several PRs when possible. This has the benefit of making code changes clearer, making PRs easier to review, and overall smoothening the whole process. No changes unrelated to the PR should be included.
When relevant, PR names should also include tags if they fall in various categories. When opening a PR, the authors should include the [WIP] tag in its name, or use github draft mode. When ready for review, they should switch the tag to [MRG] or can switch it back to normal mode. Other tags can describe the PR content : [FIX] for a bugfix, [DOC] for a change in documentation or examples, [ENH] for a new feature and [MAINT] for maintenance changes.
The nilearn codebase follows PEP8 styling. The main conventions we enforce are :
- line length < 80
- spaces around operators
- meaningful variable names
- function names are underscore separated (e.g.,
a_nice_function) and as short as possible - public functions exposed in their parent module's init file
- private function names preceded with a "_" and very explicit
- classes in CamelCase
- 2 empty lines between functions or classes
Each function and class must come with a “docstring” at the top of the function code, using numpydoc formatting. The docstring must summarize what the function does and document every parameter.
Additionally, we consider it best practice to write modular functions; i.e., functions should preferably be relatively short and do one thing. This is also useful for writing unit tests.
Writing small functions is not always possible, and we do not recommend trying to reorganize larger, but well-tested, older functions in the codebase, unless there is a strong reason to do so (e.g., when adding a new feature).
When fixing a bug, the first step is to write a minimal test that fails because of it, and then write the bugfix to make this test pass. For new code you should have roughly one test function per function covering every line and testing the logic of the function. They should run on small mocked data, cover a representative range of parameters.
Hint
It is easier to write good unit tests for short, self-contained functions. Try to keep this in mind when you write new functions. For more information about this coding approach, see test-driven development.
Tests must be seeded to avoid random failures. For objects using random seeds (e.g. scikit-learn estimators), pass either a np.random.RandomState or an int as the seed. When your test use random numbers, those must be generated through:
rng = np.random.RandomState(0)
my_number = rng.normal()To check your changes worked and didn't break anything run pytest nilearn. To do quicker checks it's possible to run only a subset of tests:
pytest -v test_module.py
Documentation must be understandable by people from different backgrounds. The “narrative” documentation should be an introduction to the concepts of the library. It includes very little code and should first help the user figure out which parts of the library he needs and then how to use it. It must be full of links, of easily-understandable titles, colorful boxes and figures.
Examples take a hands-on approach focused on a generic usecase from which users will be able to adapt code to solve their own problems. They include plain text for explanations, python code and its output and most importantly figures to depict its results. Each example should take only a few seconds to run.
To build our documentation, we are using sphinx for the main documentation and sphinx-gallery for the example tutorials. If you want to work on those, check out next section to learn how to use those tools to build documentation.
Here are the key steps you need to go through to copy the repo before contributing:
fork the repo from github (fork button in the top right corner of our :nilearn-gh:`main github page <>`) and clone your fork locally:
git clone git@github.com:<your_username>/nilearn.git
(optional but highly recommended) set up a conda environment to work on and activate it:
conda create -n nilearn conda activate nilearn
install the forked version of nilearn:
pip install -e '.[dev]'
This installs your local version of Nilearn, along with all dependencies necessary for developers (hence the [dev] tag).
For more information about the dependency installation options, see setup.cfg.
The installed version will also reflect any changes you make to your code.
check that all tests pass with (this can take a while):
pytest nilearn
Here are the key steps you need to go through to contribute code to nilearn:
open or join an already existing issue explaining what you want to work on
on your fork, create a new branch from main:
git checkout -b your_branch
implement changes and (optional but highly recommended) lint:
Recommendation
To lint your code and verify PEP8 compliance, you can run flake8 locally on the changes you have made in your branch compared to the main branch. To do this, find the latest common ancestor (commit) of your branch with main and then get the diff between your working directory and this commit and pipe it to flake8 by running:
COMMIT=$(git merge-base main @)
git diff $COMMIT | flake8 --diffcommit your changes on this branch (don't forget to write tests!)
run the tests locally (to go faster, only run tests which are relevant to what you work on with, for example):
pytest -v nilearn/plotting/tests/test_surf_plotting.py
push your changes to your online fork:
git push
in github, open a pull request from your online fork to the main repo (most likely from your_fork:your_branch to nilearn:main).
check that all continuous integration tests pass
For more details about the Fork Clone Push workflows, read here.
If you wish to build documentation:
First, ensure that you have installed sphinx and sphinx-gallery. When in your fork top folder, you can install the required packages using:
pip install '.[doc]'
Then go to
nilearn/examplesornilearn/docand make needed changes using reStructuredText filesYou can now go to nilearn/doc and build the examples locally:
make html-strict
or, if you do not have make install (for instance under Windows):
python3 -m sphinx -b html -d _build/doctrees . _build/html
if you don't need the plots, a quicker option is:
make html-noplot
Visually review the output in
nilearn/doc/_build/html/auto_examples/. If all looks well and there were no errors, commit and push the changes.You can now open a Pull Request from Nilearn's Pull Request page.
Request the CI builds the full documentation from your branch:
git commit --allow-empty -m "[circle full] request full build"
Tip
When generating documentation locally, you can build only specific files
to reduce building time. To do so, use the filename_pattern:
python3 -m sphinx -D sphinx_gallery_conf.filename_pattern=\\ plot_decoding_tutorial.py -b html -d _build/doctrees . _build/html
We want atlases in nilearn to be internally consistent. Specifically, your atlas object should have three attributes (as with the existing atlases):
description(bytes): A text description of the atlas. This should be brief but thorough, describing the source (paper), relevant information related to its construction (modality, dataset, method), and, if there is more than one map, a description of each map.labels(list): a list of string labels corresponding to each atlas label, in the same (numerical) order as the atlas labelsmaps(list or string): the path to the nifti image, or a list of paths
In addition, the atlas will need to be called by a fetcher. For example, see :nilearn-gh:`here <blob/main/nilearn/datasets/atlas.py>`.
Finally, as with other features, please provide a test for your atlas. Examples can be found :nilearn-gh:`here <blob/main/nilearn/datasets/tests/test_atlas.py>`.
The :mod:`nilearn.datasets` module provides functions to download some neuroimaging datasets, such as :func:`nilearn.datasets.fetch_haxby` or :func:`nilearn.datasets.fetch_atlas_harvard_oxford`. The goal is not to provide a comprehensive collection of downloaders for the most widely used datasets, and this would be outside the scope of this project. Rather, this module provides data downloading utilities that are required to showcase nilearn features in the example gallery.
Downloading data takes time and large datasets slow down the build of the example gallery. Moreover, downloads can fail for reasons we do not control, such as a web service that is temporarily unavailable. This is frustrating for users and a major issue for continuous integration (new code cannot be merged unless the examples run successfully on the CI infrastructure). Finally, datasets or the APIs that provide them sometimes change, in which case the downloader needs to be adapted.
As for any contributed feature, before starting working on a new downloader, we recommend opening a :nilearn-gh:`new issue <issues/new/choose>` to discuss whether it is necessary or if existing downloaders could be used instead.
To add a new fetcher, nilearn.datasets.utils provides some helper functions,
such as _get_dataset_dir to find a directory where the dataset is or will be
stored according to the user's configuration, or _fetch_files to load files
from the disk or download them if they are missing.
The new fetcher, as any other function, also needs to be tested (in the relevant
submodule of nilearn.datasets.tests). When the tests run, the fetcher does
not have access to the network and will not actually download files. This is to
avoid spurious failures due to unavailable network or servers, and to avoid
slowing down the tests with long downloads.
The functions from the standard library and the requests library that
nilearn uses to download files are mocked: they are replaced with dummy
functions that return fake data.
Exactly what fake data is returned can be configured through the object
returned by the request_mocker pytest fixture, defined in
nilearn.datasets._testing. The docstrings of this module and the Sender
class it contains provide information on how to write a test using this fixture.
Existing tests can also serve as examples.
More information about the project organization, conventions, and maintenance process can be found there : :ref:`maintenance_process`.