This project is a community effort, and everyone is welcome to contribute.

The project is hosted on

The decision making process and governance structure of scikit-learn is laid out in the governance document: Scikit-learn governance and decision-making.

Scikit-learn is somewhat selective when it comes to adding new algorithms, and the best way to contribute and to help the project is to start working on known issues. See Issues for New Contributors to get started.

Our community, our values

We are a community based on openness and friendly, didactic, discussions.

We aspire to treat everybody equally, and value their contributions.

Decisions are made based on technical merit and consensus.

Code is not the only way to help the project. Reviewing pull requests, answering questions to help others on mailing lists or issues, organizing and teaching tutorials, working on the website, improving the documentation, are all priceless contributions.

We abide by the principles of openness, respect, and consideration of others of the Python Software Foundation:

In case you experience issues using this package, do not hesitate to submit a ticket to the GitHub issue tracker. You are also welcome to post feature requests or pull requests.

Ways to contribute

There are many ways to contribute to scikit-learn, with the most common ones being contribution of code or documentation to the project. Improving the documentation is no less important than improving the library itself. If you find a typo in the documentation, or have made improvements, do not hesitate to send an email to the mailing list or preferably submit a GitHub pull request. Full documentation can be found under the doc/ directory.

But there are many other ways to help. In particular answering queries on the issue tracker, investigating bugs, and reviewing other developers’ pull requests are very valuable contributions that decrease the burden on the project maintainers.

Another way to contribute is to report issues you’re facing, and give a “thumbs up” on issues that others reported and that are relevant to you. It also helps us if you spread the word: reference the project from your blog and articles, link to it from your website, or simply star to say “I use it”:

In case a contribution/issue involves changes to the API principles or changes to dependencies or supported versions, it must be backed by a Enhancement proposals (SLEPs), where a SLEP must be submitted as a pull-request to enhancement proposals using the SLEP template and follows the decision-making process outlined in Scikit-learn governance and decision-making.


Contributing to related projects

Scikit-learn thrives in an ecosystem of several related projects, which also may have relevant issues to work on, including smaller projects such as:

and larger projects:

Look for issues marked “help wanted” or similar. Helping these projects may help Scikit-learn too. See also Related Projects.

Submitting a bug report or a feature request

We use GitHub issues to track all bugs and feature requests; feel free to open an issue if you have found a bug or wish to see a feature implemented.

In case you experience issues using this package, do not hesitate to submit a ticket to the Bug Tracker. You are also welcome to post feature requests or pull requests.

It is recommended to check that your issue complies with the following rules before submitting:

How to make a good bug report

When you submit an issue to Github, please do your best to follow these guidelines! This will make it a lot easier to provide you with good feedback:

  • The ideal bug report contains a short reproducible code snippet, this way anyone can try to reproduce the bug easily (see this for more details). If your snippet is longer than around 50 lines, please link to a gist or a github repo.

  • If not feasible to include a reproducible snippet, please be specific about what estimators and/or functions are involved and the shape of the data.

  • If an exception is raised, please provide the full traceback.

  • Please include your operating system type and version number, as well as your Python, scikit-learn, numpy, and scipy versions. This information can be found by running the following code snippet:

    >>> import sklearn
    >>> sklearn.show_versions()  


    This utility function is only available in scikit-learn v0.20+. For previous versions, one has to explicitly run:

    import platform; print(platform.platform())
    import sys; print("Python", sys.version)
    import numpy; print("NumPy", numpy.__version__)
    import scipy; print("SciPy", scipy.__version__)
    import sklearn; print("Scikit-Learn", sklearn.__version__)
  • Please ensure all code snippets and error messages are formatted in appropriate code blocks. See Creating and highlighting code blocks for more details.

Contributing code


To avoid duplicating work, it is highly advised that you search through the issue tracker and the PR list. If in doubt about duplicated work, or if you want to work on a non-trivial feature, it’s recommended to first open an issue in the issue tracker to get some feedbacks from core developers.

How to contribute

The preferred way to contribute to scikit-learn is to fork the main repository on GitHub, then submit a “pull request” (PR):

  1. Create an account on GitHub if you do not already have one.

  2. Fork the project repository: click on the ‘Fork’ button near the top of the page. This creates a copy of the code under your account on the GitHub user account. For more details on how to fork a repository see this guide.

  3. Clone your fork of the scikit-learn repo from your GitHub account to your local disk:

    $ git clone
    $ cd scikit-learn
  4. Install the development dependencies:

    $ pip install cython pytest flake8
  5. Install scikit-learn in editable mode:

    $ pip install --editable .

    for more details about advanced installation, see the Building from source section.

  6. Add the upstream remote. This saves a reference to the main scikit-learn repository, which you can use to keep your repository synchronized with the latest changes:

    $ git remote add upstream
  7. Fetch the upstream and then create a branch to hold your development changes:

    $ git fetch upstream
    $ git checkout -b my-feature upstream/master

    and start making changes. Always use a feature branch. It’s good practice to never work on the master branch!

  8. Develop the feature on your feature branch on your computer, using Git to do the version control. When you’re done editing, add changed files using git add and then git commit files:

    $ git add modified_files
    $ git commit

    to record your changes in Git, then push the changes to your GitHub account with:

    $ git push -u origin my-feature
  9. Follow these instructions to create a pull request from your fork. This will send an email to the committers. You may want to consider sending an email to the mailing list for more visibility.


If you are modifying a Cython module, you have to re-run step 5 after modifications and before testing them.

It is often helpful to keep your local branch synchronized with the latest changes of the main scikit-learn repository:

$ git fetch upstream
$ git merge upstream/master

Subsequently, you might need to solve the conflicts. You can refer to the Git documentation related to resolving merge conflict using the command line.

Learning git:

The Git documentation and are excellent resources to get started with git, and understanding all of the commands shown here.

Pull request checklist

Before a PR can be merged, it needs to be approved by two core developers. Please prefix the title of your pull request with [MRG] if the contribution is complete and should be subjected to a detailed review. An incomplete contribution – where you expect to do more work before receiving a full review – should be prefixed [WIP] (to indicate a work in progress) and changed to [MRG] when it matures. WIPs may be useful to: indicate you are working on something to avoid duplicated work, request broad review of functionality or API, or seek collaborators. WIPs often benefit from the inclusion of a task list in the PR description.

In order to ease the reviewing process, we recommend that your contribution complies with the following rules before marking a PR as [MRG]. The bolded ones are especially important:

  1. Give your pull request a helpful title that summarises what your contribution does. This title will often become the commit message once merged so it should summarise your contribution for posterity. In some cases “Fix <ISSUE TITLE>” is enough. “Fix #<ISSUE NUMBER>” is never a good title.

  2. Make sure your code passes the tests. The whole test suite can be run with pytest, but it is usually not recommended since it takes a long time. It is often enough to only run the test related to your changes: for example, if you changed something in sklearn/linear_model/, running the following commands will usually be enough:

    • pytest sklearn/linear_model/ to make sure the doctest examples are correct
    • pytest sklearn/linear_model/tests/ to run the tests specific to the file
    • pytest sklearn/linear_model to test the whole Generalized Linear Models module
    • pytest sklearn/doc/linear_model.rst to make sure the user guide examples are correct.
    • pytest sklearn/tests/ -k LogisticRegression to run all our estimator checks (specifically for LogisticRegression, if that’s the estimator you changed).

    There may be other failing tests, but they will be caught by the CI so you don’t need to run the whole test suite locally. You can read more in Testing and improving test coverage.

  3. Make sure your code is properly commented and documented, and make sure the documentation renders properly. To build the documentation, please refer to our Documentation guidelines. The CI will also build the docs: please refer to Generated documentation on CircleCI.

  4. Tests are necessary for enhancements to be accepted. Bug-fixes or new features should be provided with non-regression tests. These tests verify the correct behavior of the fix or feature. In this manner, further modifications on the code base are granted to be consistent with the desired behavior. In the case of bug fixes, at the time of the PR, the non-regression tests should fail for the code base in the master branch and pass for the PR code.

  5. Make sure that your PR does not add PEP8 violations. On a Unix-like system, you can run make flake8-diff. flake8 path_to_file, would work for any system, but please avoid reformatting parts of the file that your pull request doesn’t change, as it distracts from code review.

  6. Follow the coding-guidelines (see below).

  7. When applicable, use the validation tools and scripts in the sklearn.utils submodule. A list of utility routines available for developers can be found in the Utilities for Developers page.

  8. Often pull requests resolve one or more other issues (or pull requests). If merging your pull request means that some other issues/PRs should be closed, you should use keywords to create link to them (e.g., Fixes #1234; multiple issues/PRs are allowed as long as each one is preceded by a keyword). Upon merging, those issues/PRs will automatically be closed by GitHub. If your pull request is simply related to some other issues/PRs, create a link to them without using the keywords (e.g., See also #1234).

  9. PRs should often substantiate the change, through benchmarks of performance and efficiency or through examples of usage. Examples also illustrate the features and intricacies of the library to users. Have a look at other examples in the examples/ directory for reference. Examples should demonstrate why the new functionality is useful in practice and, if possible, compare it to other methods available in scikit-learn.

  10. New features often need to be illustrated with narrative documentation in the user guide, with small code snipets. If relevant, please also add references in the literature, with PDF links when possible.

  11. The user guide should also include expected time and space complexity of the algorithm and scalability, e.g. “this algorithm can scale to a large number of samples > 100000, but does not scale in dimensionality: n_features is expected to be lower than 100”.

You can also check our Code Review Guidelines to get an idea of what reviewers will expect.

You can check for common programming errors with the following tools:

  • Code with a good unittest coverage (at least 80%, better 100%), check with:

    $ pip install pytest pytest-cov
    $ pytest --cov sklearn path/to/tests_for_package

    see also Testing and improving test coverage

Bonus points for contributions that include a performance analysis with a benchmark script and profiling output (please report on the mailing list or on the GitHub issue).

Also check out the How to optimize for speed guide for more details on profiling and Cython optimizations.


The current state of the scikit-learn code base is not compliant with all of those guidelines, but we expect that enforcing those constraints on all new contributions will get the overall code base quality in the right direction.


For two very well documented and more detailed guides on development workflow, please pay a visit to the Scipy Development Workflow - and the Astropy Workflow for Developers sections.

Continuous Integration (CI)

  • Azure pipelines are used for testing scikit-learn on Linux, Mac and Windows, with different dependencies and settings.
  • CircleCI is used to build the docs for viewing, for linting with flake8, and for testing with PyPy on Linux

Please note that if one of the following markers appear in the latest commit message, the following actions are taken.

Commit Message Marker Action Taken by CI
[scipy-dev] Add a Travis build with our dependencies (numpy, scipy, etc …) development builds
[ci skip] CI is skipped completely
[doc skip] Docs are not built
[doc quick] Docs built, but excludes example gallery plots
[doc build] Docs built including example gallery plots

Stalled pull requests

As contributing a feature can be a lengthy process, some pull requests appear inactive but unfinished. In such a case, taking them over is a great service for the project.

A good etiquette to take over is:

  • Determine if a PR is stalled

    • A pull request may have the label “stalled” or “help wanted” if we have already identified it as a candidate for other contributors.

    • To decide whether an inactive PR is stalled, ask the contributor if she/he plans to continue working on the PR in the near future. Failure to respond within 2 weeks with an activity that moves the PR forward suggests that the PR is stalled and will result in tagging that PR with “help wanted”.

      Note that if a PR has received earlier comments on the contribution that have had no reply in a month, it is safe to assume that the PR is stalled and to shorten the wait time to one day.

      After a sprint, follow-up for un-merged PRs opened during sprint will be communicated to participants at the sprint, and those PRs will be tagged “sprint”. PRs tagged with “sprint” can be reassigned or declared stalled by sprint leaders.

  • Taking over a stalled PR: To take over a PR, it is important to comment on the stalled PR that you are taking over and to link from the new PR to the old one. The new PR should be created by pulling from the old one.

Issues for New Contributors

New contributors should look for the following tags when looking for issues. We strongly recommend that new contributors tackle “easy” issues first: this helps the contributor become familiar with the contribution workflow, and for the core devs to become acquainted with the contributor; besides which, we frequently underestimate how easy an issue is to solve!

good first issue tag

A great way to start contributing to scikit-learn is to pick an item from the list of good first issues in the issue tracker. Resolving these issues allow you to start contributing to the project without much prior knowledge. If you have already contributed to scikit-learn, you should look at Easy issues instead.

Easy tag

If you have already contributed to scikit-learn, another great way to contribute to scikit-learn is to pick an item from the list of Easy issues in the issue tracker. Your assistance in this area will be greatly appreciated by the more experienced developers as it helps free up their time to concentrate on other issues.

help wanted tag

We often use the help wanted tag to mark issues regardless of difficulty. Additionally, we use the help wanted tag to mark Pull Requests which have been abandoned by their original contributor and are available for someone to pick up where the original contributor left off. The list of issues with the help wanted tag can be found here .

Note that not all issues which need contributors will have this tag.


We are glad to accept any sort of documentation: function docstrings, reStructuredText documents (like this one), tutorials, etc. reStructuredText documents live in the source code repository under the doc/ directory.

You can edit the documentation using any text editor, and then generate the HTML output by typing make from the doc/ directory. Alternatively, make html may be used to generate the documentation with the example gallery (which takes quite some time). The resulting HTML files will be placed in _build/html/stable and are viewable in a web browser.

Building the documentation

First, make sure you have properly installed the development version.

Building the documentation requires installing some additional packages:

pip install sphinx sphinx-gallery numpydoc matplotlib Pillow pandas scikit-image

To build the documentation, you need to be in the doc folder:

cd doc

In the vast majority of cases, you only need to generate the full web site, without the example gallery:


The documentation will be generated in the _build/html/stable directory. To also generate the example gallery you can use:

make html

This will run all the examples, which takes a while. If you only want to generate a few examples, you can use:

EXAMPLES_PATTERN=your_regex_goes_here make html

This is particularly useful if you are modifying a few examples.

Set the environment variable NO_MATHJAX=1 if you intend to view the documentation in an offline setting.

To build the PDF manual, run:

make latexpdf


Sphinx version

While we do our best to have the documentation build under as many versions of Sphinx as possible, the different versions tend to behave slightly differently. To get the best results, you should use the same version as the one we used on CircleCI. Look at this github search to know the exact version.

Guidelines for writing documentation

It is important to keep a good compromise between mathematical and algorithmic details, and give intuition to the reader on what the algorithm does.

Basically, to elaborate on the above, it is best to always start with a small paragraph with a hand-waving explanation of what the method does to the data. Then, it is very helpful to point out why the feature is useful and when it should be used - the latter also including “big O” (\(O\left(g\left(n\right)\right)\)) complexities of the algorithm, as opposed to just rules of thumb, as the latter can be very machine-dependent. If those complexities are not available, then rules of thumb may be provided instead.

Secondly, a generated figure from an example (as mentioned in the previous paragraph) should then be included to further provide some intuition.

Next, one or two small code examples to show its use can be added.

Next, any math and equations, followed by references, can be added to further the documentation. Not starting the documentation with the maths makes it more friendly towards users that are just interested in what the feature will do, as opposed to how it works “under the hood”.

Finally, follow the formatting rules below to make it consistently good:

  • Add “See also” in docstrings for related classes/functions.

  • “See also” in docstrings should be one line per reference, with a colon and an explanation, for example:

    See also
    SelectKBest : Select features based on the k highest scores.
    SelectFpr : Select features based on a false positive rate test.
  • For unwritten formatting rules, try to follow existing good works:

  • When editing reStructuredText (.rst) files, try to keep line length under 80 characters when possible (exceptions include links and tables).

Generated documentation on CircleCI

When you change the documentation in a pull request, CircleCI automatically builds it. To view the documentation generated by CircleCI:

  • navigate to the bottom of your pull request page to see the CI statuses. You may need to click on “Show all checks” to see all the CI statuses.
  • click on the CircleCI status with “doc” in the title.
  • add #artifacts at the end of the URL. Note: you need to wait for the CircleCI build to finish before being able to look at the artifacts.
  • once the artifacts are visible, navigate to doc/_changed.html to see a list of documentation pages that are likely to be affected by your pull request. Navigate to doc/index.html to see the full generated html documentation.

If you often need to look at the documentation generated by CircleCI, e.g. when reviewing pull requests, you may find this tip very handy.

Testing and improving test coverage

High-quality unit testing is a corner-stone of the scikit-learn development process. For this purpose, we use the pytest package. The tests are functions appropriately named, located in tests subdirectories, that check the validity of the algorithms and the different options of the code.

The full scikit-learn tests can be run using ‘make’ in the root folder. Alternatively, running ‘pytest’ in a folder will run all the tests of the corresponding subpackages.

We expect code coverage of new features to be at least around 90%.

For guidelines on how to use pytest efficiently, see the Useful pytest aliases and flags.

Workflow to improve test coverage

To test code coverage, you need to install the coverage package in addition to pytest.

  1. Run ‘make test-coverage’. The output lists for each file the line
    numbers that are not tested.
  2. Find a low hanging fruit, looking at which lines are not tested,
    write or adapt a test specifically for these lines.
  3. Loop.

Issue Tracker Tags

All issues and pull requests on the GitHub issue tracker should have (at least) one of the following tags:

Bug / Crash:Something is happening that clearly shouldn’t happen. Wrong results as well as unexpected errors from estimators go here.
Cleanup / Enhancement:
 Improving performance, usability, consistency.
Documentation:Missing, incorrect or sub-standard documentations and examples.
New Feature:Feature requests and pull requests implementing a new feature.

There are four other tags to help new contributors:

good first issue:
 This issue is ideal for a first contribution to scikit-learn. Ask for help if the formulation is unclear. If you have already contributed to scikit-learn, look at Easy issues instead.
Easy:This issue can be tackled without much prior experience.
Moderate:Might need some knowledge of machine learning or the package, but is still approachable for someone new to the project.
help wanted:This tag marks an issue which currently lacks a contributor or a PR that needs another contributor to take over the work. These issues can range in difficulty, and may not be approachable for new contributors. Note that not all issues which need contributors will have this tag.

Coding guidelines

The following are some guidelines on how new code should be written. Of course, there are special cases and there will be exceptions to these rules. However, following these rules when submitting new code makes the review easier so new code can be integrated in less time.

Uniformly formatted code makes it easier to share code ownership. The scikit-learn project tries to closely follow the official Python guidelines detailed in PEP8 that detail how code should be formatted and indented. Please read it and follow it.

In addition, we add the following guidelines:

  • Use underscores to separate words in non class names: n_samples rather than nsamples.
  • Avoid multiple statements on one line. Prefer a line return after a control flow statement (if/for).
  • Use relative imports for references inside scikit-learn.
  • Unit tests are an exception to the previous rule; they should use absolute imports, exactly as client code would. A corollary is that, if exports a class or function that is implemented in, the test should import it from
  • Please don’t use import * in any case. It is considered harmful by the official Python recommendations. It makes the code harder to read as the origin of symbols is no longer explicitly referenced, but most important, it prevents using a static analysis tool like pyflakes to automatically find bugs in scikit-learn.
  • Use the numpy docstring standard in all your docstrings.

A good example of code that we like can be found here.

Input validation

The module sklearn.utils contains various functions for doing input validation and conversion. Sometimes, np.asarray suffices for validation; do not use np.asanyarray or np.atleast_2d, since those let NumPy’s np.matrix through, which has a different API (e.g., * means dot product on np.matrix, but Hadamard product on np.ndarray).

In other cases, be sure to call check_array on any array-like argument passed to a scikit-learn API function. The exact parameters to use depends mainly on whether and which scipy.sparse matrices must be accepted.

For more information, refer to the Utilities for Developers page.

Random Numbers

If your code depends on a random number generator, do not use numpy.random.random() or similar routines. To ensure repeatability in error checking, the routine should accept a keyword random_state and use this to construct a numpy.random.RandomState object. See sklearn.utils.check_random_state in Utilities for Developers.

Here’s a simple example of code using some of the above guidelines:

from sklearn.utils import check_array, check_random_state

def choose_random_sample(X, random_state=0):
    Choose a random point from X

    X : array-like, shape (n_samples, n_features)
        array representing the data
    random_state : RandomState or an int seed (0 by default)
        A random number generator instance to define the state of the
        random permutations generator.

    x : numpy array, shape (n_features,)
        A random point selected from X
    X = check_array(X)
    random_state = check_random_state(random_state)
    i = random_state.randint(X.shape[0])
    return X[i]

If you use randomness in an estimator instead of a freestanding function, some additional guidelines apply.

First off, the estimator should take a random_state argument to its __init__ with a default value of None. It should store that argument’s value, unmodified, in an attribute random_state. fit can call check_random_state on that attribute to get an actual random number generator. If, for some reason, randomness is needed after fit, the RNG should be stored in an attribute random_state_. The following example should make this clear:

class GaussianNoise(BaseEstimator, TransformerMixin):
    """This estimator ignores its input and returns random Gaussian noise.

    It also does not adhere to all scikit-learn conventions,
    but showcases how to handle randomness.

    def __init__(self, n_components=100, random_state=None):
        self.random_state = random_state

    # the arguments are ignored anyway, so we make them optional
    def fit(self, X=None, y=None):
        self.random_state_ = check_random_state(self.random_state)

    def transform(self, X):
        n_samples = X.shape[0]
        return self.random_state_.randn(n_samples, n_components)

The reason for this setup is reproducibility: when an estimator is fit twice to the same data, it should produce an identical model both times, hence the validation in fit, not __init__.


If any publicly accessible method, function, attribute or parameter is renamed, we still support the old one for two releases and issue a deprecation warning when it is called/passed/accessed. E.g., if the function zero_one is renamed to zero_one_loss, we add the decorator deprecated (from sklearn.utils) to zero_one and call zero_one_loss from that function:

from ..utils import deprecated

def zero_one_loss(y_true, y_pred, normalize=True):
    # actual implementation

@deprecated("Function 'zero_one' was renamed to 'zero_one_loss' "
            "in version 0.13 and will be removed in release 0.15. "
            "Default behavior is changed from 'normalize=False' to "
def zero_one(y_true, y_pred, normalize=False):
    return zero_one_loss(y_true, y_pred, normalize)

If an attribute is to be deprecated, use the decorator deprecated on a property. Please note that the property decorator should be placed before the deprecated decorator for the docstrings to be rendered properly. E.g., renaming an attribute labels_ to classes_ can be done as:

@deprecated("Attribute labels_ was deprecated in version 0.13 and "
            "will be removed in 0.15. Use 'classes_' instead")
def labels_(self):
    return self.classes_

If a parameter has to be deprecated, use DeprecationWarning appropriately. In the following example, k is deprecated and renamed to n_clusters:

import warnings

def example_function(n_clusters=8, k='not_used'):
    if k != 'not_used':
        warnings.warn("'k' was renamed to n_clusters in version 0.13 and "
                      "will be removed in 0.15.", DeprecationWarning)
        n_clusters = k

When the change is in a class, we validate and raise warning in fit:

import warnings

class ExampleEstimator(BaseEstimator):
    def __init__(self, n_clusters=8, k='not_used'):
        self.n_clusters = n_clusters
        self.k = k

    def fit(self, X, y):
        if self.k != 'not_used':
            warnings.warn("'k' was renamed to n_clusters in version 0.13 and "
                          "will be removed in 0.15.", DeprecationWarning)
            self._n_clusters = self.k
            self._n_clusters = self.n_clusters

As in these examples, the warning message should always give both the version in which the deprecation happened and the version in which the old behavior will be removed. If the deprecation happened in version 0.x-dev, the message should say deprecation occurred in version 0.x and the removal will be in 0.(x+2), so that users will have enough time to adapt their code to the new behaviour. For example, if the deprecation happened in version 0.18-dev, the message should say it happened in version 0.18 and the old behavior will be removed in version 0.20.

In addition, a deprecation note should be added in the docstring, recalling the same information as the deprecation warning as explained above. Use the .. deprecated:: directive:

.. deprecated:: 0.13
   ``k`` was renamed to ``n_clusters`` in version 0.13 and will be removed
   in 0.15.

What’s more, a deprecation requires a test which ensures that the warning is raised in relevant cases but not in other cases. The warning should be caught in all other tests (using e.g., @pytest.mark.filterwarnings), and there should be no warning in the examples.

Change the default value of a parameter

If the default value of a parameter needs to be changed, please replace the default value with a specific value (e.g., warn) and raise FutureWarning when users are using the default value. In the following example, we change the default value of n_clusters from 5 to 10 (current version is 0.20):

import warnings

def example_function(n_clusters='warn'):
    if n_clusters == 'warn':
        warnings.warn("The default value of n_clusters will change from "
                      "5 to 10 in 0.22.", FutureWarning)
        n_clusters = 5

When the change is in a class, we validate and raise warning in fit:

import warnings

class ExampleEstimator:
    def __init__(self, n_clusters='warn'):
        self.n_clusters = n_clusters

    def fit(self, X, y):
        if self.n_clusters == 'warn':
          warnings.warn("The default value of n_clusters will change from "
                        "5 to 10 in 0.22.", FutureWarning)
          self._n_clusters = 5

Similar to deprecations, the warning message should always give both the version in which the change happened and the version in which the old behavior will be removed. The docstring needs to be updated accordingly. We need a test which ensures that the warning is raised in relevant cases but not in other cases. The warning should be caught in all other tests (using e.g., @pytest.mark.filterwarnings), and there should be no warning in the examples.

Python versions supported

Since scikit-learn 0.21, only Python 3.5 and newer is supported.

Code Review Guidelines

Reviewing code contributed to the project as PRs is a crucial component of scikit-learn development. We encourage anyone to start reviewing code of other developers. The code review process is often highly educational for everybody involved. This is particularly appropriate if it is a feature you would like to use, and so can respond critically about whether the PR meets your needs. While each pull request needs to be signed off by two core developers, you can speed up this process by providing your feedback.

Here are a few important aspects that need to be covered in any code review, from high-level questions to a more detailed check-list.

  • Do we want this in the library? Is it likely to be used? Do you, as a scikit-learn user, like the change and intend to use it? Is it in the scope of scikit-learn? Will the cost of maintaining a new feature be worth its benefits?
  • Is the code consistent with the API of scikit-learn? Are public functions/classes/parameters well named and intuitively designed?
  • Are all public functions/classes and their parameters, return types, and stored attributes named according to scikit-learn conventions and documented clearly?
  • Is any new functionality described in the user-guide and illustrated with examples?
  • Is every public function/class tested? Are a reasonable set of parameters, their values, value types, and combinations tested? Do the tests validate that the code is correct, i.e. doing what the documentation says it does? If the change is a bug-fix, is a non-regression test included? Look at this to get started with testing in Python.
  • Do the tests pass in the continuous integration build? If appropriate, help the contributor understand why tests failed.
  • Do the tests cover every line of code (see the coverage report in the build log)? If not, are the lines missing coverage good exceptions?
  • Is the code easy to read and low on redundancy? Should variable names be improved for clarity or consistency? Should comments be added? Should comments be removed as unhelpful or extraneous?
  • Could the code easily be rewritten to run much more efficiently for relevant settings?
  • Is the code backwards compatible with previous versions? (or is a deprecation cycle necessary?)
  • Will the new code add any dependencies on other libraries? (this is unlikely to be accepted)
  • Does the documentation render properly (see the Documentation section for more details), and are the plots instructive?

Standard replies for reviewing includes some frequent comments that reviewers may make.

APIs of scikit-learn objects

To have a uniform API, we try to have a common basic API for all the objects. In addition, to avoid the proliferation of framework code, we try to adopt simple conventions and limit to a minimum the number of methods an object must implement.

Elements of the scikit-learn API are described more definitively in the Glossary of Common Terms and API Elements.

Different objects

The main objects in scikit-learn are (one class can implement multiple interfaces):


The base object, implements a fit method to learn from data, either:

estimator =, targets)


estimator =

For supervised learning, or some unsupervised problems, implements:

prediction = predictor.predict(data)

Classification algorithms usually also offer a way to quantify certainty of a prediction, either using decision_function or predict_proba:

probability = predictor.predict_proba(data)

For filtering or modifying the data, in a supervised or unsupervised way, implements:

new_data = transformer.transform(data)

When fitting and transforming can be performed much more efficiently together than separately, implements:

new_data = transformer.fit_transform(data)

A model that can give a goodness of fit measure or a likelihood of unseen data, implements (higher is better):

score = model.score(data)


The API has one predominant object: the estimator. A estimator is an object that fits a model based on some training data and is capable of inferring some properties on new data. It can be, for instance, a classifier or a regressor. All estimators implement the fit method:, y)

All built-in estimators also have a set_params method, which sets data-independent parameters (overriding previous parameter values passed to __init__).

All estimators in the main scikit-learn codebase should inherit from sklearn.base.BaseEstimator.


This concerns the creation of an object. The object’s __init__ method might accept constants as arguments that determine the estimator’s behavior (like the C constant in SVMs). It should not, however, take the actual training data as an argument, as this is left to the fit() method:

clf2 = SVC(C=2.3)
clf3 = SVC([[1, 2], [2, 3]], [-1, 1]) # WRONG!

The arguments accepted by __init__ should all be keyword arguments with a default value. In other words, a user should be able to instantiate an estimator without passing any arguments to it. The arguments should all correspond to hyperparameters describing the model or the optimisation problem the estimator tries to solve. These initial arguments (or parameters) are always remembered by the estimator. Also note that they should not be documented under the “Attributes” section, but rather under the “Parameters” section for that estimator.

In addition, every keyword argument accepted by __init__ should correspond to an attribute on the instance. Scikit-learn relies on this to find the relevant attributes to set on an estimator when doing model selection.

To summarize, an __init__ should look like:

def __init__(self, param1=1, param2=2):
    self.param1 = param1
    self.param2 = param2

There should be no logic, not even input validation, and the parameters should not be changed. The corresponding logic should be put where the parameters are used, typically in fit. The following is wrong:

def __init__(self, param1=1, param2=2, param3=3):
    # WRONG: parameters should not be modified
    if param1 > 1:
        param2 += 1
    self.param1 = param1
    # WRONG: the object's attributes should have exactly the name of
    # the argument in the constructor
    self.param3 = param2

The reason for postponing the validation is that the same validation would have to be performed in set_params, which is used in algorithms like GridSearchCV.


The next thing you will probably want to do is to estimate some parameters in the model. This is implemented in the fit() method.

The fit() method takes the training data as arguments, which can be one array in the case of unsupervised learning, or two arrays in the case of supervised learning.

Note that the model is fitted using X and y, but the object holds no reference to X and y. There are, however, some exceptions to this, as in the case of precomputed kernels where this data must be stored for use by the predict method.

X array-like, shape (n_samples, n_features)
y array, shape (n_samples,)
kwargs optional data-dependent parameters.

X.shape[0] should be the same as y.shape[0]. If this requisite is not met, an exception of type ValueError should be raised.

y might be ignored in the case of unsupervised learning. However, to make it possible to use the estimator as part of a pipeline that can mix both supervised and unsupervised transformers, even unsupervised estimators need to accept a y=None keyword argument in the second position that is just ignored by the estimator. For the same reason, fit_predict, fit_transform, score and partial_fit methods need to accept a y argument in the second place if they are implemented.

The method should return the object (self). This pattern is useful to be able to implement quick one liners in an IPython session such as:

y_predicted = SVC(C=100).fit(X_train, y_train).predict(X_test)

Depending on the nature of the algorithm, fit can sometimes also accept additional keywords arguments. However, any parameter that can have a value assigned prior to having access to the data should be an __init__ keyword argument. fit parameters should be restricted to directly data dependent variables. For instance a Gram matrix or an affinity matrix which are precomputed from the data matrix X are data dependent. A tolerance stopping criterion tol is not directly data dependent (although the optimal value according to some scoring function probably is).

When fit is called, any previous call to fit should be ignored. In general, calling and then should be the same as only calling However, this may not be true in practice when fit depends on some random process, see random_state. Another exception to this rule is when the hyper-parameter warm_start is set to True for estimators that support it. warm_start=True means that the previous state of the trainable parameters of the estimator are reused instead of using the default initialization strategy.

Estimated Attributes

Attributes that have been estimated from the data must always have a name ending with trailing underscore, for example the coefficients of some regression estimator would be stored in a coef_ attribute after fit has been called.

The estimated attributes are expected to be overridden when you call fit a second time.

Optional Arguments

In iterative algorithms, the number of iterations should be specified by an integer called n_iter.

Pairwise Attributes

An estimator that accept X of shape (n_samples, n_samples) and defines a _pairwise property equal to True allows for cross-validation of the dataset, e.g. when X is a precomputed kernel matrix. Specifically, the _pairwise property is used by utils.metaestimators._safe_split to slice rows and columns.

Rolling your own estimator

If you want to implement a new estimator that is scikit-learn-compatible, whether it is just for you or for contributing it to scikit-learn, there are several internals of scikit-learn that you should be aware of in addition to the scikit-learn API outlined above. You can check whether your estimator adheres to the scikit-learn interface and standards by running utils.estimator_checks.check_estimator on the class:

>>> from sklearn.utils.estimator_checks import check_estimator
>>> from sklearn.svm import LinearSVC
>>> check_estimator(LinearSVC)  # passes

The main motivation to make a class compatible to the scikit-learn estimator interface might be that you want to use it together with model evaluation and selection tools such as model_selection.GridSearchCV and pipeline.Pipeline.

Before detailing the required interface below, we describe two ways to achieve the correct interface more easily.

Project template:

We provide a project template which helps in the creation of Python packages containing scikit-learn compatible estimators. It provides:

  • an initial git repository with Python package directory structure
  • a template of a scikit-learn estimator
  • an initial test suite including use of check_estimator
  • directory structures and scripts to compile documentation and example galleries
  • scripts to manage continuous integration (testing on Linux and Windows)
  • instructions from getting started to publishing on PyPi

BaseEstimator and mixins:

We tend to use “duck typing”, so building an estimator which follows the API suffices for compatibility, without needing to inherit from or even import any scikit-learn classes.

However, if a dependency on scikit-learn is acceptable in your code, you can prevent a lot of boilerplate code by deriving a class from BaseEstimator and optionally the mixin classes in sklearn.base. For example, below is a custom classifier, with more examples included in the scikit-learn-contrib project template.

>>> import numpy as np
>>> from sklearn.base import BaseEstimator, ClassifierMixin
>>> from sklearn.utils.validation import check_X_y, check_array, check_is_fitted
>>> from sklearn.utils.multiclass import unique_labels
>>> from sklearn.metrics import euclidean_distances
>>> class TemplateClassifier(BaseEstimator, ClassifierMixin):
...     def __init__(self, demo_param='demo'):
...         self.demo_param = demo_param
...     def fit(self, X, y):
...         # Check that X and y have correct shape
...         X, y = check_X_y(X, y)
...         # Store the classes seen during fit
...         self.classes_ = unique_labels(y)
...         self.X_ = X
...         self.y_ = y
...         # Return the classifier
...         return self
...     def predict(self, X):
...         # Check is fit had been called
...         check_is_fitted(self, ['X_', 'y_'])
...         # Input validation
...         X = check_array(X)
...         closest = np.argmin(euclidean_distances(X, self.X_), axis=1)
...         return self.y_[closest]

get_params and set_params

All scikit-learn estimators have get_params and set_params functions. The get_params function takes no arguments and returns a dict of the __init__ parameters of the estimator, together with their values. It must take one keyword argument, deep, which receives a boolean value that determines whether the method should return the parameters of sub-estimators (for most estimators, this can be ignored). The default value for deep should be true.

The set_params on the other hand takes as input a dict of the form 'parameter': value and sets the parameter of the estimator using this dict. Return value must be estimator itself.

While the get_params mechanism is not essential (see Cloning below), the set_params function is necessary as it is used to set parameters during grid searches.

The easiest way to implement these functions, and to get a sensible __repr__ method, is to inherit from sklearn.base.BaseEstimator. If you do not want to make your code dependent on scikit-learn, the easiest way to implement the interface is:

def get_params(self, deep=True):
    # suppose this estimator has parameters "alpha" and "recursive"
    return {"alpha": self.alpha, "recursive": self.recursive}

def set_params(self, **parameters):
    for parameter, value in parameters.items():
        setattr(self, parameter, value)
    return self

Parameters and init

As model_selection.GridSearchCV uses set_params to apply parameter setting to estimators, it is essential that calling set_params has the same effect as setting parameters using the __init__ method. The easiest and recommended way to accomplish this is to not do any parameter validation in __init__. All logic behind estimator parameters, like translating string arguments into functions, should be done in fit.

Also it is expected that parameters with trailing _ are not to be set inside the __init__ method. All and only the public attributes set by fit have a trailing _. As a result the existence of parameters with trailing _ is used to check if the estimator has been fitted.


For use with the model_selection module, an estimator must support the base.clone function to replicate an estimator. This can be done by providing a get_params method. If get_params is present, then clone(estimator) will be an instance of type(estimator) on which set_params has been called with clones of the result of estimator.get_params().

Objects that do not provide this method will be deep-copied (using the Python standard function copy.deepcopy) if safe=False is passed to clone.

Pipeline compatibility

For an estimator to be usable together with pipeline.Pipeline in any but the last step, it needs to provide a fit or fit_transform function. To be able to evaluate the pipeline on any data but the training set, it also needs to provide a transform function. There are no special requirements for the last step in a pipeline, except that it has a fit function. All fit and fit_transform functions must take arguments X, y, even if y is not used. Similarly, for score to be usable, the last step of the pipeline needs to have a score function that accepts an optional y.

Estimator types

Some common functionality depends on the kind of estimator passed. For example, cross-validation in model_selection.GridSearchCV and model_selection.cross_val_score defaults to being stratified when used on a classifier, but not otherwise. Similarly, scorers for average precision that take a continuous prediction need to call decision_function for classifiers, but predict for regressors. This distinction between classifiers and regressors is implemented using the _estimator_type attribute, which takes a string value. It should be "classifier" for classifiers and "regressor" for regressors and "clusterer" for clustering methods, to work as expected. Inheriting from ClassifierMixin, RegressorMixin or ClusterMixin will set the attribute automatically. When a meta-estimator needs to distinguish among estimator types, instead of checking _estimator_type directly, helpers like base.is_classifier should be used.

Specific models

Classifiers should accept y (target) arguments to fit that are sequences (lists, arrays) of either strings or integers. They should not assume that the class labels are a contiguous range of integers; instead, they should store a list of classes in a classes_ attribute or property. The order of class labels in this attribute should match the order in which predict_proba, predict_log_proba and decision_function return their values. The easiest way to achieve this is to put:

self.classes_, y = np.unique(y, return_inverse=True)

in fit. This returns a new y that contains class indexes, rather than labels, in the range [0, n_classes).

A classifier’s predict method should return arrays containing class labels from classes_. In a classifier that implements decision_function, this can be achieved with:

def predict(self, X):
    D = self.decision_function(X)
    return self.classes_[np.argmax(D, axis=1)]

In linear models, coefficients are stored in an array called coef_, and the independent term is stored in intercept_. sklearn.linear_model.base contains a few base classes and mixins that implement common linear model patterns.

The sklearn.utils.multiclass module contains useful functions for working with multiclass and multilabel problems.

Estimator Tags


The estimator tags are experimental and the API is subject to change.

Scikit-learn introduced estimator tags in version 0.21. These are annotations of estimators that allow programmatic inspection of their capabilities, such as sparse matrix support, supported output types and supported methods. The estimator tags are a dictionary returned by the method _get_tags(). These tags are used by the common tests and the sklearn.utils.estimator_checks.check_estimator function to decide what tests to run and what input data is appropriate. Tags can depend on estimator parameters or even system architecture and can in general only be determined at runtime.

The default value of all tags except for X_types is False. These are defined in the BaseEstimator class.

The current set of estimator tags are:

whether the estimator is not deterministic given a fixed random_state
requires_positive_data - unused for now
whether the estimator requires positive X.
whether the estimator skips input-validation. This is only meant for stateless and dummy transformers!
multioutput - unused for now
whether a regressor supports multi-target outputs or a classifier supports multi-class multi-output.
whether the estimator supports multilabel output
whether the estimator needs access to data for fitting. Even though an estimator is stateless, it might still need a call to fit for initialization.
whether the estimator supports data with missing values encoded as np.NaN
whether the estimator fails to provide a “reasonable” test-set score, which currently for regression is an R2 of 0.5 on a subset of the boston housing dataset, and for classification an accuracy of 0.83 on make_blobs(n_samples=300, random_state=0). These datasets and values are based on current estimators in sklearn and might be replaced by something more systematic.
whether estimator supports only multi-output classification or regression.
whether to skip common tests entirely. Don’t use this unless you have a very good reason.
Supported input types for X as list of strings. Tests are currently only run if ‘2darray’ is contained in the list, signifying that the estimator takes continuous 2d numpy arrays as input. The default value is [‘2darray’]. Other possible types are 'string', 'sparse', 'categorical', dict, '1dlabels' and '2dlabels'. The goal is that in the future the supported input type will determine the data used during testing, in particular for 'string', 'sparse' and 'categorical' data. For now, the test for sparse data do not make use of the 'sparse' tag.

To override the tags of a child class, one must define the _more_tags() method and return a dict with the desired tags, e.g:

class MyMultiOutputEstimator(BaseEstimator):

    def _more_tags(self):
        return {'multioutput_only': True,
                'non_deterministic': True}

In addition to the tags, estimators also need to declare any non-optional parameters to __init__ in the _required_parameters class attribute, which is a list or tuple. If _required_parameters is only ["estimator"] or ["base_estimator"], then the estimator will be instantiated with an instance of LinearDiscriminantAnalysis (or RidgeRegression if the estimator is a regressor) in the tests. The choice of these two models is somewhat idiosyncratic but both should provide robust closed-form solutions.

Reading the existing code base

Reading and digesting an existing code base is always a difficult exercise that takes time and experience to master. Even though we try to write simple code in general, understanding the code can seem overwhelming at first, given the sheer size of the project. Here is a list of tips that may help make this task easier and faster (in no particular order).

  • Get acquainted with the APIs of scikit-learn objects: understand what fit, predict, transform, etc. are used for.
  • Before diving into reading the code of a function / class, go through the docstrings first and try to get an idea of what each parameter / attribute is doing. It may also help to stop a minute and think how would I do this myself if I had to?
  • The trickiest thing is often to identify which portions of the code are relevant, and which are not. In scikit-learn a lot of input checking is performed, especially at the beginning of the fit methods. Sometimes, only a very small portion of the code is doing the actual job. For example looking at the fit() method of sklearn.linear_model.LinearRegression, what you’re looking for might just be the call the scipy.linalg.lstsq, but it is buried into multiple lines of input checking and the handling of different kinds of parameters.
  • Due to the use of Inheritance, some methods may be implemented in parent classes. All estimators inherit at least from BaseEstimator, and from a Mixin class (e.g. ClassifierMixin) that enables default behaviour depending on the nature of the estimator (classifier, regressor, transformer, etc.).
  • Sometimes, reading the tests for a given function will give you an idea of what its intended purpose is. You can use git grep (see below) to find all the tests written for a function. Most tests for a specific function/class are placed under the tests/ folder of the module
  • You’ll often see code looking like this: out = Parallel(...)(delayed(some_function)(param) for param in some_iterable). This runs some_function in parallel using Joblib. out is then an iterable containing the values returned by some_function for each call.
  • We use Cython to write fast code. Cython code is located in .pyx and .pxd files. Cython code has a more C-like flavor: we use pointers, perform manual memory allocation, etc. Having some minimal experience in C / C++ is pretty much mandatory here.
  • Master your tools.
    • With such a big project, being efficient with your favorite editor or IDE goes a long way towards digesting the code base. Being able to quickly jump (or peek) to a function/class/attribute definition helps a lot. So does being able to quickly see where a given name is used in a file.
    • git also has some built-in killer features. It is often useful to understand how a file changed over time, using e.g. git blame (manual). This can also be done directly on GitHub. git grep (examples) is also extremely useful to see every occurrence of a pattern (e.g. a function call or a variable) in the code base.