updated: 2019-05-11: This post mostly mentions
virtualenv, but according to the Python doc about module installation, since Python 3.5 "the use of
venv is now recommended for creating virtual environments", while
virtualenv is an alternative for versions of Python prior to 3.4.
updated: 2018-08-17: since conda-4.4.0 use
activate anaconda on all platforms
updated: 2017-03-27: PEP 513 -
manylinux binaries for PyPI
updated: 2016-08-19: Continuum Anaconda Option
This is somewhat a duplicate of easy_install/pip or apt-get.
For global Python packages, use either the Ubuntu Software Center, apt, apt-get or synaptic
Ubuntu uses Python for many important functions, therefore interfering with Python can corrupt your OS. This is the main reason I never use
pip on my Ubuntu system, but instead I use either Ubuntu Software Center, synaptic,
apt-get, or the newer just
apt, which all by default install packages from the Ubuntu repository. These packages are tested, usually pre-compiled so they install faster and ultimately designed for Ubuntu. In addition all required dependencies are also installed and a log of installs is maintained so they can be rolled back. I think most packages have corresponding Launchpad repos so you can file issues.
Another reason to use either Ubuntu packages is that sometimes these Python packages have different names depending on where you downloaded them from. Python-chardet is an example of a package which at one time was named one thing on PyPI and another thing in the Ubuntu repository. Therefore doing something like
pip install requests will not realize that chardet is already installed in your system because the Ubuntu version has a different name, and consequently install a new version which will corrupt your system in a minor insignificant way but still why would you do that.
In general you only want to install trusted code into your OS. So you should be nervous about typing
$ sudo pip <anything-could-be-very-bad>.
Lastly some things are just easier to install using either Ubuntu packages. For example if you try
pip install numpy to install numpy & scipy unless you have already installed gfortran, atlas-dev, blas-dev and lapack-dev, you will see an endless stream of compile errors. However, installing numpy & scipy through the Ubuntu repository is as easy as...
$ sudo apt-get install python-numpy python-scipy
You are in luck, because you are using Ubuntu, one of the most widely supported and oft updated distributions existing. Most likely every Python package you will need is in the Ubuntu repository, and probably already installed on your machine. And every 6 months, a new cycle of packages will be released with the latest distribution of Ubuntu.
If you are 100% confident that the package will not interfere with your Ubuntu system in any way, then you can install it using pip and Ubuntu is nice enough to keep these packages separate from the distro packages by placing the distro packages in a folder called
dist-packages/. Ubuntu repository has both pip, virtualenv and setuptools. However, I second Wojciech's suggestion to use virtualenv.
For personal Python projects use pip and wheel in a virtualenv
If you need the latest version, or the module is not in the Ubuntu repository then start a virtualenv and use pip to install the package. Although pip and setuptools have merged, IMO pip is preferred over easy-install or distutils, because it will always wait until the package is completely downloaded and built before it copies it into your file system, and it makes upgrading or uninstalling a breeze. In a lot of ways it is similar to apt-get, in that it generally handles dependencies well. However you
will may have to handle some dependencies yourself, but since PEP 513 was adopted there are now
manylinux binaries at the Python Package Index (PyPI) for popular Linux distros like Ubuntu and Fedora.
for example as mentioned above for NumPy and SciPy make sure you have installed gfortran, atlas-dev, blas-dev and lapack-dev from the Ubuntu repository For example, both NumPy and SciPy are now distributed for Ubuntu as
manylinux wheels by default using OpenBLAS instead of ATLAS. You can still build them from source by using the pip options
--no-binary <format control>.
~$ sudo apt-get install gfortran libblas-dev liblapack-dev libatlas-dev python-virtualenv
~$ mkdir ~/.venvs
~$ virtualenv ~/.venvs/my_py_proj
~$ source ~/.venvs/my_py_proj/bin/activate
~(my_py_proj)$ pip install --no-use-wheel numpy scipy
See the next section, "You're not in
sudoers", below for installing updated versions of pip, setuptools, virtualenv or wheels to your personal profile using the
--user installation scheme with pip. You can use this to update pip for your personal use as J.F. Sebastian indicated in his comment to another answer. NOTE: the
-m is really only necessary on MS Windows when updating pip.
python -m pip install --user pip setuptools wheel virtualenv
Newer versions of pip automatically cache wheels, so the following is only useful for older versions of pip. Since you may end up installing these many times, consider using wheel with pip to create a wheelhouse. Wheel is already included in
virtualenv since v13.0.0 therefore if your version of
virtualenv is too old, you may need to install wheel first.
~(my_py_proj)$ pip install wheel # only for virtualenv < v13.0.0
~(my_py_proj)$ pip wheel --no-use-wheel numpy scipy
This will create binary wheel files in
-d to specify a different directory. Now if you start another virtualenv and you need the same packages you've already built, you can install them form your wheelhouse using
pip install --find-links=<fullpath>/wheelhouse
Read Installing Python Modules in the Python documentation and Installing packages on the Python Package Index main page. Also pip, venv, virtualenv and wheel.
If you're not in
virtualenv isn't installed.
Another option to using a virtual environment, or if you are using a Linux share without root privileges, then using either the
--home=<wherever-you-want> Python installation schemes with Python's
distutils will install packages to the value of
site.USERBASE or to wherever you want. Newer versions of pip also have a
--user option. Do not use
pip install --user virtualenv
If your Linux version of pip is too old, then you can pass setup options using
--install-option which is useful for passing custom options to some
setup.py scripts for some packages that build extensions, such as setting the
PREFIX. You may need to just extract the distribution and use
distutils to install the package the old-school way by typing
python setup install [options]. Reading some of the install documentation and the
distutils documentation may help.
Python is nice enough to add
site.USERBASE to your
PYTHONPATH ahead of anything else, so the changes will only effect you. A popular location for
~/.local. See the Python module installation guide for the exact file structure and specifically where your site-packages are. Note: if you use the
--home installation scheme then you may need to add it to the
PYTHONPATH environment variable using
export in your
.bash_profile or in your shell for your localized packages to be available in Python.
If you are using Python for either math, science, or data, then IMO a really good option is the Anaconda-Python Distribution or the more basic miniconda distro released by Anaconda, Inc. (previously known as Continuum Analytics). Although anyone could benefit from using Anaconda for personal projects, the default installation includes over 500 math and science packages like NumPy, SciPy, Pandas, and Matplotlib, while miniconda only installs Anaconda-Python and the conda environment manager. Anaconda only installs into your personal profile, ie:
alters your recommends sourcing
~/.bash_profile to prepend Anaconda's path to your personal
conda.sh in your
~/.bashrc which lets you use
conda activate <env|default is base> to start anaconda - this only affects you - your system path is unchanged. Therefore you do not need root access or
sudo to use Anaconda! If you have already added Anaconda-Python, miniconda, or conda to your personal path, then you should remove the
PATH export from your
~/.bashrc, and update to the new recommendation, so your system Python will be first again.
This is somewhat similar to the
--user option I explained in the last section except it applies to Python as a whole and not just packages. Therefore Anaconda is completely separate from your system Python, it won't interfere with your system Python, and only you can use or change it. Since it installs a new version of Python and all of its libraries you will need at least 200MB of room, but it is very clever about caching and managing libraries which is important for some of the cool things you can do with Anaconda.
Anaconda curates a collection of Python binaries and libraries required by dependencies in an online repository (formerly called binstar), and they also host user packages as different "channels". The package manager used by Anaconda,
conda, by default installs packages from Anaconda, but you can signal a different "channel" using the
Install packages with
conda just like
$ conda install -c pvlib pvlib # install pvlib pkg from pvlib channel
conda can do so much more! It can also create and manage virtual environments just like
virtualenv. Therefore since Anaconda creates virtual environments, the
pip package manager can be used to install packages from PyPI into an Anaconda environment without root or
sudo. Do not use
sudo with Anaconda! Warning! Do be careful though when mixing
conda in an Anaconda environment, b/c you will have to manage package dependencies more carefully. Another option to
pip in a conda environment is to use the conda-forge channel, but also best to do that in a fresh conda environment with conda-forge as the default channel. As a last resort, if you can't find a package anywhere but on PyPI, consider using
--no-deps then install the remaining dependencies manually using
Anaconda is also similar in some ways to Ruby RVM if you're familiar with that tool. Anaconda
conda also lets you create virtual environments with different versions of Python. e.g.:
conda create -n py35sci python==3.5.2 numpy scipy matplotlib pandas statsmodels seaborn will create a scientific/data-science stack using Python-3.5 in a new environment called
py35sci. You can switch environments using
conda. Since conda-4.4.0, this is now different to
virtualenv which uses
source venv/bin/activate, but previous to conda-4.4.0 the
conda commands were the same as
virtualenv and also used
# AFTER conda-4.4
~/Projects/myproj $ conda activate py35sci
# BEFORE conda-4.4
~/Projects/myproj $ source activate py35sci
But wait there's more! Anaconda can also install different languages such as R for statistical programming from the Anaconda
r channel. You can even set up your own channel to upload package distributions built for conda. As mentioned conda-forge maintains automated builds of many of the packages on PyPI at the conda-forge Anaconda channel.
There are many options for maintaining your Python projects on Linux depending on your personal needs and access. However, if there's any one thing I hope you take away from this answer is that you should almost never need to use
sudo to install Python packages. The use of
sudo should be a smell to you that something is amiss. You have been warned.
Good luck and happy coding!