UPDATE: You should probably follow the official Python Packaging user guide instead of this guide.
This tutorial will teach you to manage a project, and publish it on PyPI. This guide is majorly influenced by the following tutorial.
Also, this tutorial will always be a work in progress (or at least so long as best practice can change), so the tutorial might change at any time. However, you can always read old versions of the tutorial, since it is covered by source control. Finally, if you have any constructive critic on the contents in this tutorial, please raise an Issue with the Issue tracker.
Contents
An integral part of having reusable code is having a sensible repository structure. That is, which files do we have and how do we organise them. Unfortunately, figuring out how to structure a Python project best is not a trivial task. In this part of the tutorial, I hope to show you a way to initate any Python project to ensure that you won't have to do major effort restructuring the code once you want to publish it.
Let us start with the folder layout. Your project directory should be structured in the following way and we will explain why later.
project_name
├── docs
│ ├── make.bat
│ ├── Makefile
│ └── source
│ ├── conf.py
│ └── index.rst
├── examples
│ └── example.py
├── src
│ └── package_name
│ └── __init__.py
├── tests
│ └── __init__.py
├── .gitignore
├── LICENSE.txt
├── MANIFEST.in
├── README.rst
├── requirements.txt
├── setup.cfg
├── setup.py
└── tox.ini
Now, this is a lot of files, let us look at these to understand what the different components are and why they are necessary in a Python project.
The setup.py
, setup.cfg
and MANIFEST.in
files are used to
specify how a package should be installed. You might think that you don't
want to create an installable package, so let's skip this. DON'T! Even for
small projects, you should include these because of something called
editable installs (more on that later). The most basic setup.py file should
look like this
from setuptools import setup
setup()
Some projects might include more code, especially if you are using Cython
or creating C-extensions to Python. However, if you are not, then this style
will probably suffice. The reason we keep the setup.py
minimal is that
we want to keep as much of the setup configuration as possible inside the
setup.cfg
file. This is to let other people parse metadata about our
package without running a Python file first! The setup.cfg
file should
look like this
[metadata]
name = <package-name>
version = <version number: 0.0.0>
license = <license name, e.g. MIT>
description = <A short description>
long_description = file: README.rst
author = <Author name>
author_email = <Optional: author e-mail>
classifiers=
<classifier 1>
<classifier 2>
<...>
<classifier m>
[options]
packages = find:
package_dir =
=src
include_package_data = True
install_requires =
<requirement 1>
<requirement 2>
<...>
<requirement n>
[options.packages.find]
where=src
This file is formated according to this
specification. However, if you you
simply follow the layout above, replacing the elements wrapped in <>
with
the correct information for your package, then you are ok.
There are two sections here that might be confusing, the classifiers
section and the install_requires
section. The classifiers
section is
used by PyPI to make it easier for new users to find your package, you can find a full list of classifiers here. Likewise, the
install_requires
section specifies which Python packages that pip
should
install before installing the package you are developing. Both these fields are
optional, so you can leave them blank until you have anything to fill in.
Lastly, the MANIFEST.in
file. This file is used to instruct setupttools
on which files it should include when it creates an installable project. For
a general project, I reccomend having a file with the following layout.
include setup.py
include MANIFEST.in
include LICENSE
include README.md
graft tests
graft examples
graft docs
graft src
The requirements.txt
file is similar to the install_requires
field in
the setup.cfg
file we described above. However, the aim of the
requirements.txt
file is not to specify the dependencies of your package,
but the packages needed to work on developing your package. Each dependency
should be on a separate line. Here is an example of a requirements.txt
file.
scikit-learn
tox
black
isort
-e .
We will depend on scikit-learn
if we are to create scikit-learn compliant
code. Similarly, we need tox
to run our test-suite. black
and isort
are two really good code auto-formatters, which you can read more about on
their GitHub pages (black and isort). Finally, with the -e .
line
we install the current directory in editable mode.
The readme file contains the contens that are showed by default on online source control providers such as GitHub, GitLab and BitBucket. Normally, this is formatted as a Markdown file. However, I reccomend that you use reStructuredText (rst) instead, since that is the file-format used by Sphinx, the most commonly used auto-documentation tool for Python.
Additionally, PyPI will only host rst formatted help strings, not Markdown
formatted ones. Thus, if you wish to make your library public for pip
installation in the future, then you should use rst to avoid writing the
same text twice.
The rst documentation is available here, and a good summary is available here.
The .gitignore
file contains instructions to Git, informing it of which
files it should not track. Examples of such files are the __pycache__
files
and IDE configuration files. You can either copy the .gitignore
file in this
repository, which should work for a large array of development environments, or
create your own .gitignore
using gitignore.io.
Your project needs an open source license, otherwise, noone will be able to use your project. I like the MIT license, which is a very open license. To decide a license, i reccomend choosealicense if you are unsure as to which license to use.
You should unit test your code. Otherwise there will be bugs, no matter how
simple the codebase is. The tool I like to use for unit testing is called
tox, and works by creating new virtual environments for each python version
you want to test the codebase with. It then installs all libraries necessary
to run the test suite before running it. These specifications are given in the
tox.ini
file, which can have the following structure
[tox]
envlist =
py35
py36
[testenv]
deps =
pytest
pytest-cov
pytest-randomly
commands =
pytest --cov=<package_name> --randomly-seed=1
The envlist
field specifies which python versions to run the code with,
the deps
field specifies the test dependencies (which might be different
from the devloper dependencies) and commands
specifies which commands to
be ran to run the test suite.
Note that tox
by itself doesn't play nice with conda
. Thus, if you
have an Anaconda or Miniconda installation of Python, then you should manually
install tox-conda
through pip
.
You might have noticed that the source files are kept inside a separate src
folder. The reason is that we should be certain that the code we are testing
is the installable code. To accomplish this, it is neccessary to structure the
code this way. For more information on this topic, see this page.
For the same reason as we keep the package source in the src folder, we keep the unit tests in the tests folder.
When you publish code, you should also publish documentation to that code, and creating the documentation is very simple if you have good docstrings and use sphinx. To use sphinx, navigate to the docs folder in the terminal window and type sphinx-quickstart.
We will not discuss sphinx in detail here, the only extra note I want to add is to use the sphinx.ext.napoleon extension so your docstrings can be in the numpydoc style.
Any library should come with at least a minimal example script so prospective users can see how the package was intended to be used. Keep these example scripts in the examples folder.
One immensely useful facet of the python ecosystem is editable installs. Often, when new Python programmers create a project, they do not install the project with pip. Rather, whenever they need to use the code from one project within another, they end up manually modifying the system path environment variable. If this sounds familiar, then you should stop that immediately. There is a cleaner, easier and less error-prone way to accomplish the same. This way is called editable installs.
Normally when we install a Python package, it is copied into the
site-packages
directory. This is not ideal if the code we installed
is code that we are actively developing. In this case, we want to create a
symbolic link between the site-packages
directory and the project
directory, and a way to accomplish this is through editable installs.
To installl a project in editable mode, simply navigate to the project root
directory and type pip install -e .
in the terminal window. A benefit of
doing it this way, is that we have better cross-platform support. Windows and
UNIX based systems have vastly different ways of handling the path variable, so
your old sys.path.append
hack might not work as intended on a Windows
machine. Additionally, the sys.path.append
method is highly dependent on the
file-structure on your computer, whereas editable installs are not.
The second most important part of a project, after the source code itself, is the documentation. Luckily, writing Python documentation is relatively painless so long as you write your docstrings following the Sphinx guidelines. I will assume that you have a working sphinx environment and simply want to host the documentation somewhere.
If you are in this category, then you are in luck since you can host your
documentation for free on Read the Docs. To do this, you need to connect your GitHub
user to https://readthedocs.org (note the org top level domain (TLD), not
an io TLD). Once you have connected your GitHub to Read the Docs, you need
to add the .readthedocs.yml
file to your repository. This file should have
the following lines in it.
python:
setup_py_install: true
After adding the .readthedocs.yml
file to the repository, it should have
the following layout.
project_name
├── docs
│ ├── make.bat
│ ├── Makefile
│ └── source
│ ├── conf.py
│ └── index.rst
├── examples
│ └── example.py
├── src
│ └── package_name
│ └── __init__.py
├── tests
│ └── test_package_name
│ └── __init__.py
├── .gitignore
├── .readthedocs.yml <- This file is new
├── LICENSE.txt
├── MANIFEST.in
├── README.rst
├── requirements.txt
├── setup.cfg
├── setup.py
└── tox.ini
Once it does, you can import the project to Read the Docs, by pressing the "Import a Project" button and choosing the correct GitHub repository.
You might want to have a badge that shows whether your documentation builds correctly on your GitHub page, to do this, press the "i" button on the right of the green "docs passing" badge (or red "docs failing" if your documentation isn't building correctly). Copy the rst code to somewhere near the beginning of your readme file. The code should be on the following form:
.. image:: https://readthedocs.org/projects/<repo_name>/badge/?version=latest
:target: https://<repo_name>.readthedocs.io/en/latest/?badge=latest
:alt: Documentation Status
Another useful tool when developing code is a continuous integration tool. Such tools will automatically run the unit tests on activity to the GitHub repository. Luckily, there exists a very good tool called *Travis-CI*, which is free for all open source projects.
To use Travis-CI, you must link your GitHub user to Travis CI on their webpage.
After this, you simply choose which repository to activate Travis for and you
are set to go. When you have activated Travis for a specific repo, you need
to add a .travis.yml
file to the project root, giving you the following
file structure
project_name
├── docs
│ ├── make.bat
│ ├── Makefile
│ └── source
│ ├── conf.py
│ └── index.rst
├── examples
│ └── example.py
├── src
│ └── package_name
│ └── __init__.py
├── tests
│ └── test_package_name
│ └── __init__.py
├── .gitignore
├── .readthedocs.yml
├── .travis.yml <- This file is new
├── LICENSE.txt
├── MANIFEST.in
├── README.rst
├── requirements.txt
├── setup.cfg
├── setup.py
└── tox.ini
The contents of the .travis.yml
file should be the following
sudo: false
language: python
python:
- "3.7"
# command to install dependencies
install:
before_script:
- pip install tox-travis
# command to run tests
script: tox
This file will ensure that tox is run on Travis-CI any time someone pushes
a change to the GitHub repository. You might also want to add a badge to
your readme file. To do this, navigate to the Travis-CI dashboard, press
the link to the repository that you want to add the badge for, press the
badge showing build passing
(ideally, it will show build failing
if your tests are failing) and finally, choose rst from the bottom dropdown
menu. Once you have done this, copy the text in the text-box and paste it
somewhere around the top of yor README.rst
file. The rst code that you
copy should look something like this
.. image:: https://travis-ci.org/<github_username>/<repo_name>.svg?branch=<branch_name>
:target: https://travis-ci.org/<github_username>/<repo_name>
Another useful tool in a programmer's arsenal is automatic code coverage reporting. Have you ever seen a repository where they have a badge that shows how high their code-coverage is with a small badge? They accomplish this using one of many automatic code-coverage reporters. Personally, I like to use *Coveralls*, which has a relatively easy-to-use interface and integrates well with Travis-CI.
To start using Coveralls, you must first register and link your GitHub account
with Coveralls. Once you have done that, you need to add your repository to
Coveralls. You can do this, by pressing the plus button on the left-hand side of
the Coveralls dashboard and enable whichever repository you want. Once you have
done this, you must update the .travis.yml
file so Coveralls are ran after
the test suite. The new .travis.yml
file should look like this:
sudo: false
language: python
python:
- "3.7"
# command to install dependencies
install:
before_script:
- pip install tox-travis
- pip install coveralls
# command to run tests
script: tox
after_success: coveralls
Once you have made this update, then Coveralls will run after travis. Next, you
want to add the coverage badge to your README.rst
file. In the Coveralls
project dashboard, you should see a badge that displays your code coverage,
press the embed button on the top right corner near the badge and copy the
code for rst into your README.rst
file. The code you copy should have
the following format
.. image:: https://coveralls.io/repos/github/<github_username>/<repo_name>/badge.svg?branch=<branch_name>
:target: https://coveralls.io/github/<github_username>/<repo_name>?branch=<branch_name>
It is finally time to upload our code to PyPI, making it easily installable for
others. Uploading code to PyPI is very simple. First, create an account on PyPI.
Then, you need to install two packages; twine and wheel. To do this, write
pip install twine wheel
in the terminal window. Then, navigate to the
project root and type python setup.py sdist bdist_wheel
, this will prepare
your package for uploading to PyPI. Then, write twine upload dist/*
to
upload your project.