Installing the development version of scikit-learn

This section introduces how to install the main branch of scikit-learn. This can be done by either installing a nightly build or building from source.

Installing nightly builds

The continuous integration servers of the scikit-learn project build, test and upload wheel packages for the most recent Python version on a nightly basis.

Installing a nightly build is the quickest way to:

  • try a new feature that will be shipped in the next release (that is, a feature from a pull-request that was recently merged to the main branch);

  • check whether a bug you encountered has been fixed since the last release.

You can install the nightly build of scikit-learn using the scientific-python-nightly-wheels index from the PyPI registry of anaconda.org:

pip install --pre --extra-index https://pypi.anaconda.org/scientific-python-nightly-wheels/simple scikit-learn

Note that first uninstalling scikit-learn might be required to be able to install nightly builds of scikit-learn.

Building from source

Building from source is required to work on a contribution (bug fix, new feature, code or documentation improvement).

  1. Use Git to check out the latest source from the scikit-learn repository on Github.:

    git clone git://github.com/scikit-learn/scikit-learn.git  # add --depth 1 if your connection is slow
    cd scikit-learn
    

    If you plan on submitting a pull-request, you should clone from your fork instead.

  2. Install a recent version of Python (3.9 is recommended at the time of writing) for instance using Miniforge3. Miniforge provides a conda-based distribution of Python and the most popular scientific libraries.

    If you installed Python with conda, we recommend to create a dedicated conda environment with all the build dependencies of scikit-learn (namely NumPy, SciPy, and Cython):

    conda create -n sklearn-env -c conda-forge python=3.9 numpy scipy cython
    

    It is not always necessary but it is safer to open a new prompt before activating the newly created conda environment.

    conda activate sklearn-env
    
  3. Alternative to conda: You can use alternative installations of Python provided they are recent enough (3.9 or higher at the time of writing). Here is an example on how to create a build environment for a Linux system’s Python. Build dependencies are installed with pip in a dedicated virtualenv to avoid disrupting other Python programs installed on the system:

    python3 -m venv sklearn-env
    source sklearn-env/bin/activate
    pip install wheel numpy scipy cython
    
  4. Install a compiler with OpenMP support for your platform. See instructions for Windows, macOS, Linux and FreeBSD.

  5. Build the project with pip in Editable mode:

    pip install -v --no-use-pep517 --no-build-isolation -e .
    
  6. Check that the installed scikit-learn has a version number ending with .dev0:

    python -c "import sklearn; sklearn.show_versions()"
    
  7. Please refer to the Developer’s Guide and Useful pytest aliases and flags to run the tests on the module of your choice.

Note

You will have to run the pip install -v --no-use-pep517 --no-build-isolation -e . command every time the source code of a Cython file is updated (ending in .pyx or .pxd). This can happen when you edit them or when you use certain git commands such as git pull. Use the --no-build-isolation flag to avoid compiling the whole project each time, only the files you have modified. Include the --no-use-pep517 flag because the --no-build-isolation option might not work otherwise (this is due to a bug which will be fixed in the future).

Dependencies

Runtime dependencies

Scikit-learn requires the following dependencies both at build time and at runtime:

  • Python (>= 3.8),

  • NumPy (>= 1.19.5),

  • SciPy (>= 1.6.0),

  • Joblib (>= 1.2.0),

  • threadpoolctl (>= 2.0.0).

Build dependencies

Building Scikit-learn also requires:

Note

If OpenMP is not supported by the compiler, the build will be done with OpenMP functionalities disabled. This is not recommended since it will force some estimators to run in sequential mode instead of leveraging thread-based parallelism. Setting the SKLEARN_FAIL_NO_OPENMP environment variable (before cythonization) will force the build to fail if OpenMP is not supported.

Since version 0.21, scikit-learn automatically detects and uses the linear algebra library used by SciPy at runtime. Scikit-learn has therefore no build dependency on BLAS/LAPACK implementations such as OpenBlas, Atlas, Blis or MKL.

Test dependencies

Running tests requires:

  • pytest >= 7.1.2

Some tests also require pandas.

Building a specific version from a tag

If you want to build a stable version, you can git checkout <VERSION> to get the code for that particular version, or download an zip archive of the version from github.

Editable mode

If you run the development version, it is cumbersome to reinstall the package each time you update the sources. Therefore it is recommended that you install in with the pip install -v --no-use-pep517 --no-build-isolation -e . command, which allows you to edit the code in-place. This builds the extension in place and creates a link to the development directory (see the pip docs).

As the doc above explains, this is fundamentally similar to using the command python setup.py develop. (see the setuptool docs). It is however preferred to use pip.

On Unix-like systems, you can equivalently type make in from the top-level folder. Have a look at the Makefile for additional utilities.

Building with Meson

Support for Meson is experimental, in scikit-learn 1.5.0.dev0. Open an issue if you encounter any problems!

Make sure you have meson-python and ninja installed, either with conda:

conda install -c conda-forge meson-python ninja -y

or with pip:

pip install meson-python ninja

Simplest way to build with Meson

To build scikit-learn, the simplest way is to run:

make dev-meson

You need to do it once after this you can run your code that imports sklearn and it will recompile as needed.

In case you want to go back to using setuptools:

make clean-meson

More advanced way to build with Meson

If you can not use make, want to do it yourself or understand what goes in behind the scenes, you can build scikit-learn with the following pip command:

pip install --editable . \
    --verbose --no-build-isolation \
    --config-settings editable-verbose=true

If you want to go back to using setuptools:

pip uninstall -y scikit-learn

Note --config-settings editable-verbose=true is advised to avoid surprises. meson-python implements editable install by recompiling when doing import sklearn. Even changing python files involves copying files to the Meson build directory. You will see the meson output when that happens, rather than potentially waiting a while and wondering what is taking so long. Bonus: that means you only have to do the pip install once, after that your code will recompile when doing import sklearn.

Other places that may be worth looking at:

Platform-specific instructions

Here are instructions to install a working C/C++ compiler with OpenMP support to build scikit-learn Cython extensions for each supported platform.

Windows

First, download the Build Tools for Visual Studio 2019 installer.

Run the downloaded vs_buildtools.exe file, during the installation you will need to make sure you select “Desktop development with C++”, similarly to this screenshot:

../_images/visual-studio-build-tools-selection.png

Secondly, find out if you are running 64-bit or 32-bit Python. The building command depends on the architecture of the Python interpreter. You can check the architecture by running the following in cmd or powershell console:

python -c "import struct; print(struct.calcsize('P') * 8)"

For 64-bit Python, configure the build environment by running the following commands in cmd or an Anaconda Prompt (if you use Anaconda):

SET DISTUTILS_USE_SDK=1
"C:\Program Files (x86)\Microsoft Visual Studio\2019\BuildTools\VC\Auxiliary\Build\vcvarsall.bat" x64

Replace x64 by x86 to build for 32-bit Python.

Please be aware that the path above might be different from user to user. The aim is to point to the “vcvarsall.bat” file that will set the necessary environment variables in the current command prompt.

Finally, build scikit-learn from this command prompt:

pip install -v --no-use-pep517 --no-build-isolation -e .

macOS

The default C compiler on macOS, Apple clang (confusingly aliased as /usr/bin/gcc), does not directly support OpenMP. We present two alternatives to enable OpenMP support:

  • either install conda-forge::compilers with conda;

  • or install libomp with Homebrew to extend the default Apple clang compiler.

For Apple Silicon M1 hardware, only the conda-forge method below is known to work at the time of writing (January 2021). You can install the macos/arm64 distribution of conda using the miniforge installer

macOS compilers from conda-forge

If you use the conda package manager (version >= 4.7), you can install the compilers meta-package from the conda-forge channel, which provides OpenMP-enabled C/C++ compilers based on the llvm toolchain.

First install the macOS command line tools:

xcode-select --install

It is recommended to use a dedicated conda environment to build scikit-learn from source:

conda create -n sklearn-dev -c conda-forge python numpy scipy cython \
    joblib threadpoolctl pytest compilers llvm-openmp

It is not always necessary but it is safer to open a new prompt before activating the newly created conda environment.

conda activate sklearn-dev
make clean
pip install -v --no-use-pep517 --no-build-isolation -e .

Note

If you get any conflicting dependency error message, try commenting out any custom conda configuration in the $HOME/.condarc file. In particular the channel_priority: strict directive is known to cause problems for this setup.

You can check that the custom compilers are properly installed from conda forge using the following command:

conda list

which should include compilers and llvm-openmp.

The compilers meta-package will automatically set custom environment variables:

echo $CC
echo $CXX
echo $CFLAGS
echo $CXXFLAGS
echo $LDFLAGS

They point to files and folders from your sklearn-dev conda environment (in particular in the bin/, include/ and lib/ subfolders). For instance -L/path/to/conda/envs/sklearn-dev/lib should appear in LDFLAGS.

In the log, you should see the compiled extension being built with the clang and clang++ compilers installed by conda with the -fopenmp command line flag.

macOS compilers from Homebrew

Another solution is to enable OpenMP support for the clang compiler shipped by default on macOS.

First install the macOS command line tools:

xcode-select --install

Install the Homebrew package manager for macOS.

Install the LLVM OpenMP library:

brew install libomp

Set the following environment variables:

export CC=/usr/bin/clang
export CXX=/usr/bin/clang++
export CPPFLAGS="$CPPFLAGS -Xpreprocessor -fopenmp"
export CFLAGS="$CFLAGS -I/usr/local/opt/libomp/include"
export CXXFLAGS="$CXXFLAGS -I/usr/local/opt/libomp/include"
export LDFLAGS="$LDFLAGS -Wl,-rpath,/usr/local/opt/libomp/lib -L/usr/local/opt/libomp/lib -lomp"

Finally, build scikit-learn in verbose mode (to check for the presence of the -fopenmp flag in the compiler commands):

make clean
pip install -v --no-use-pep517 --no-build-isolation -e .

Linux

Linux compilers from the system

Installing scikit-learn from source without using conda requires you to have installed the scikit-learn Python development headers and a working C/C++ compiler with OpenMP support (typically the GCC toolchain).

Install build dependencies for Debian-based operating systems, e.g. Ubuntu:

sudo apt-get install build-essential python3-dev python3-pip

then proceed as usual:

pip3 install cython
pip3 install --verbose --editable .

Cython and the pre-compiled wheels for the runtime dependencies (numpy, scipy and joblib) should automatically be installed in $HOME/.local/lib/pythonX.Y/site-packages. Alternatively you can run the above commands from a virtualenv or a conda environment to get full isolation from the Python packages installed via the system packager. When using an isolated environment, pip3 should be replaced by pip in the above commands.

When precompiled wheels of the runtime dependencies are not available for your architecture (e.g. ARM), you can install the system versions:

sudo apt-get install cython3 python3-numpy python3-scipy

On Red Hat and clones (e.g. CentOS), install the dependencies using:

sudo yum -y install gcc gcc-c++ python3-devel numpy scipy

Linux compilers from conda-forge

Alternatively, install a recent version of the GNU C Compiler toolchain (GCC) in the user folder using conda:

conda create -n sklearn-dev -c conda-forge python numpy scipy cython \
    joblib threadpoolctl pytest compilers

It is not always necessary but it is safer to open a new prompt before activating the newly created conda environment.

conda activate sklearn-dev
pip install -v --no-use-pep517 --no-build-isolation -e .

FreeBSD

The clang compiler included in FreeBSD 12.0 and 11.2 base systems does not include OpenMP support. You need to install the openmp library from packages (or ports):

sudo pkg install openmp

This will install header files in /usr/local/include and libs in /usr/local/lib. Since these directories are not searched by default, you can set the environment variables to these locations:

export CFLAGS="$CFLAGS -I/usr/local/include"
export CXXFLAGS="$CXXFLAGS -I/usr/local/include"
export LDFLAGS="$LDFLAGS -Wl,-rpath,/usr/local/lib -L/usr/local/lib -lomp"

Finally, build the package using the standard command:

pip install -v --no-use-pep517 --no-build-isolation -e .

For the upcoming FreeBSD 12.1 and 11.3 versions, OpenMP will be included in the base system and these steps will not be necessary.

Alternative compilers

The following command will build scikit-learn using your default C/C++ compiler.

pip install --verbose --editable .

If you want to build scikit-learn with another compiler handled by setuptools, use the following command:

python setup.py build_ext --compiler=<compiler> -i build_clib --compiler=<compiler>

To see the list of available compilers run:

python setup.py build_ext --help-compiler

If your compiler is not listed here, you can specify it through some environment variables (does not work on windows). This section of the setuptools documentation explains in details which environment variables are used by setuptools, and at which stage of the compilation, to set the compiler and linker options.

When setting these environment variables, it is advised to first check their sysconfig counterparts variables and adapt them to your compiler. For instance:

import sysconfig
print(sysconfig.get_config_var('CC'))
print(sysconfig.get_config_var('LDFLAGS'))

In addition, since Scikit-learn uses OpenMP, you need to include the appropriate OpenMP flag of your compiler into the CFLAGS and CPPFLAGS environment variables.

Parallel builds

It is possible to build scikit-learn compiled extensions in parallel by setting and environment variable as follows before calling the pip install or python setup.py build_ext commands:

export SKLEARN_BUILD_PARALLEL=3
pip install -v --no-use-pep517 --no-build-isolation -e .

On a machine with 2 CPU cores, it can be beneficial to use a parallelism level of 3 to overlap IO bound tasks (reading and writing files on disk) with CPU bound tasks (actually compiling).