I have a problem with how pip/setuptools is installing my package. When installing from the project directory (i.e. pip install .) my project's sub-packages are duplicated and placed in the root site-packages directory. The configuration is set entirely within pyproject.toml (with a minimal setup.py for compiling a single extension).
If my package is named mypackage which contains 3 sub-packages and depends on 3 dependencies, this is the expected directory structure under site-packages in the venv:
site-packages
- dependency1
- dependency2
- dependency3
- myproject
- subpackage1
- subpackage2
- subpackage3
Yet below is what I end up with, it looks like any folder containing any .py files are copied to root site-packages (i.e. including the venv itself and docs since they contain py files:
site-packages
- dependency1
- dependency2
- dependency3
- mypackage
- subpackage1
- subpackage2
- subpackage3
- subpackage1
- subpackage2
- subpackage3
- docs
- venv
What can I do to avoid duplicating sub-packages into the top-level site-packages directory/install correctly?
Here is my project structure:
myproject/
- pyproject.toml
- setup.py
- docs/
- myproject/
- __init__.py
- subpackage1/
- subpackage2/
- subpackage3/
- venv/
The reduced contents of pyproject.toml
[project]
name = "myproject"
requires-python = ">= 3.7"
dependencies = [
"dependency1",
"dependency2",
"dependency3",
]
[tool.setuptools]
packages = [
"myproject",
"myproject.subpackage1",
"myproject.subpackage2",
"myproject.subpackage3",
]
[build-system]
requires = ["setuptools >= 61.0.0", "cython"]
build-backend = "setuptools.build_meta"
The contents of setup.py:
from setuptools import Extension, setup
from Cython.Build import cythonize
ext_modules = [
Extension(
"subpackage1.func",
["..."],
extra_compile_args=['-fopenmp'],
extra_link_args=['-fopenmp'],
)
]
setup(ext_modules=cythonize(ext_modules))
I've just ran into the same issue.
In my case, the build directory used by pip was polluted with the "subfolders", probably because of a previous run where my package discovery settings were erroneous.
Because of this, although my configuration was (now) correct, these orphaned directories were copied to my site-packages as well.
In my case the build directory was in the folder where I called pip install . from.
If you want to find the build directory, or just check whether this is the problem, log pip's output to a file with pip install . --log foo.txt, and search for copying inside.
You should see a line like:
Arguments: ('copying', '<build directory>\\lib\\subpackage1\\bar.py', ...
Hope this helps!
When using setuptools, I can not get the installer to pull in any package_data files. Everything I've read says that the following is the correct way to do it. Can someone please advise?
setup(
name='myapp',
packages=find_packages(),
package_data={
'myapp': ['data/*.txt'],
},
include_package_data=True,
zip_safe=False,
install_requires=['distribute'],
)
where myapp/data/ is the location of the data files.
I realize that this is an old question, but for people finding their way here via Google: package_data is a low-down, dirty lie. It is only used when building binary packages (python setup.py bdist ...) but not when building source packages (python setup.py sdist ...). This is, of course, ridiculous -- one would expect that building a source distribution would result in a collection of files that could be sent to someone else to built the binary distribution.
In any case, using MANIFEST.in will work both for binary and for source distributions.
I just had this same issue. The solution, was simply to remove include_package_data=True.
After reading here, I realized that include_package_data aims to include files from version control, as opposed to merely "include package data" as the name implies. From the docs:
The data files [of include_package_data] must be under CVS or Subversion control
...
If you want finer-grained control over what files are included (for example, if
you have documentation files in your package directories and want to exclude
them from installation), then you can also use the package_data keyword.
Taking that argument out fixed it, which is coincidentally why it also worked when you switched to distutils, since it doesn't take that argument.
Following #Joe 's recommendation to remove the include_package_data=True line also worked for me.
To elaborate a bit more, I have no MANIFEST.in file. I use Git and not CVS.
Repository takes this kind of shape:
/myrepo
- .git/
- setup.py
- myproject
- __init__.py
- some_mod
- __init__.py
- animals.py
- rocks.py
- config
- __init__.py
- settings.py
- other_settings.special
- cool.huh
- other_settings.xml
- words
- __init__.py
word_set.txt
setup.py:
from setuptools import setup, find_packages
import os.path
setup (
name='myproject',
version = "4.19",
packages = find_packages(),
# package_dir={'mypkg': 'src/mypkg'}, # didnt use this.
package_data = {
# If any package contains *.txt or *.rst files, include them:
'': ['*.txt', '*.xml', '*.special', '*.huh'],
},
#
# Oddly enough, include_package_data=True prevented package_data from working.
# include_package_data=True, # Commented out.
data_files=[
# ('bitmaps', ['bm/b1.gif', 'bm/b2.gif']),
('/opt/local/myproject/etc', ['myproject/config/settings.py', 'myproject/config/other_settings.special']),
('/opt/local/myproject/etc', [os.path.join('myproject/config', 'cool.huh')]),
#
('/opt/local/myproject/etc', [os.path.join('myproject/config', 'other_settings.xml')]),
('/opt/local/myproject/data', [os.path.join('myproject/words', 'word_set.txt')]),
],
install_requires=[ 'jsonschema',
'logging', ],
entry_points = {
'console_scripts': [
# Blah...
], },
)
I run python setup.py sdist for a source distrib (haven't tried binary).
And when inside of a brand new virtual environment, I have a myproject-4.19.tar.gz, file,
and I use
(venv) pip install ~/myproject-4.19.tar.gz
...
And other than everything getting installed to my virtual environment's site-packages, those special data files get installed to /opt/local/myproject/data and /opt/local/myproject/etc.
include_package_data=True worked for me.
If you use git, remember to include setuptools-git in install_requires. Far less boring than having a Manifest or including all path in package_data ( in my case it's a django app with all kind of statics )
( pasted the comment I made, as k3-rnc mentioned it's actually helpful as is )
Using setup.cfg (setuptools ≥ 30.3.0)
Starting with setuptools 30.3.0 (released 2016-12-08), you can keep your setup.py very small and move the configuration to a setup.cfg file. With this approach, you could put your package data in an [options.package_data] section:
[options.package_data]
* = *.txt, *.rst
hello = *.msg
In this case, your setup.py can be as short as:
from setuptools import setup
setup()
For more information, see configuring setup using setup.cfg files.
There is some talk of deprecating setup.cfg in favour of pyproject.toml as proposed in PEP 518, but this is still provisional as of 2020-02-21.
Update: This answer is old and the information is no longer valid. All setup.py configs should use import setuptools. I've added a more complete answer at https://stackoverflow.com/a/49501350/64313
I solved this by switching to distutils. Looks like distribute is deprecated and/or broken.
from distutils.core import setup
setup(
name='myapp',
packages=['myapp'],
package_data={
'myapp': ['data/*.txt'],
},
)
I had the same problem for a couple of days but even this thread wasn't able to help me as everything was confusing. So I did my research and found the following solution:
Basically in this case, you should do:
from setuptools import setup
setup(
name='myapp',
packages=['myapp'],
package_dir={'myapp':'myapp'}, # the one line where all the magic happens
package_data={
'myapp': ['data/*.txt'],
},
)
The full other stackoverflow answer here
I found this post while stuck on the same problem.
My experience contradicts the experiences in the other answers.
include_package_data=True does include the data in the
bdist! The explanation in the setuptools
documentation
lacks context and troubleshooting tips, but
include_package_data works as advertised.
My setup:
Windows / Cygwin
git version 2.21.0
Python 3.8.1 Windows distribution
setuptools v47.3.1
check-manifest v0.42
Here is my how-to guide.
How-to include package data
Here is the file structure for a project I published on PyPI.
(It installs the application in __main__.py).
├── LICENSE.md
├── MANIFEST.in
├── my_package
│ ├── __init__.py
│ ├── __main__.py
│ └── _my_data <---- folder with data
│ ├── consola.ttf <---- data file
│ └── icon.png <---- data file
├── README.md
└── setup.py
Starting point
Here is a generic starting point for the setuptools.setup() in
setup.py.
setuptools.setup(
...
packages=setuptools.find_packages(),
...
)
setuptools.find_packages() includes all of my packages in the
distribution. My only package is my_package.
The sub-folder with my data, _my_data, is not considered a
package by Python because it does not contain an __init__.py,
and so find_packages() does not find it.
A solution often-cited, but incorrect, is to put an empty
__init__.py file in the _my_data folder.
This does make it a package, so it does include the folder
_my_data in the distribution. But the data files inside
_my_data are not included.
So making _my_data into a package does not help.
The solution is:
the sdist already contains the data files
add include_package_data=True to include the data files in the bdist as well
Experiment (how to test the solution)
There are three steps to make this a repeatable experiment:
$ rm -fr build/ dist/ my_package.egg-info/
$ check-manifest
$ python setup.py sdist bdist_wheel
I will break these down step-by-step:
Clean out the old build:
$ rm -fr build/ dist/ my_package.egg-info/
Run check-manifest to be sure MANIFEST.in matches the
Git index of files under version control:
$ check-manifest
If MANIFEST.in does not exist yet, create it from the Git
index of files under version control:
$ check-manifest --create
Here is the MANIFEST.in that is created:
include *.md
recursive-include my_package *.png
recursive-include my_package *.ttf
There is no reason to manually edit this file.
As long as everything that should be under version control is
under version control (i.e., is part of the Git index),
check-manifest --create does the right thing.
Note: files are not part of the Git index if they are either:
ignored in a .gitignore
excluded in a .git/info/exclude
or simply new files that have not been added to the index yet
And if any files are under version control that should not be
under version control, check-manifest issues a warning and
specifies which files it recommends removing from the Git index.
Build:
$ python setup.py sdist bdist_wheel
Now inspect the sdist (source distribution) and bdist_wheel
(build distribution) to see if they include the data files.
Look at the contents of the sdist (only the relevant lines are
shown below):
$ tar --list -f dist/my_package-0.0.1a6.tar.gz
my_package-0.0.1a6/
...
my_package-0.0.1a6/my_package/__init__.py
my_package-0.0.1a6/my_package/__main__.py
my_package-0.0.1a6/my_package/_my_data/
my_package-0.0.1a6/my_package/_my_data/consola.ttf <-- yay!
my_package-0.0.1a6/my_package/_my_data/icon.png <-- yay!
...
So the sdist already includes the data files because they are
listed in MANIFEST.in. There is nothing extra to do to include
the data files in the sdist.
Look at the contents of the bdist (it is a .zip file, parsed
with zipfile.ZipFile):
$ python check-whl.py
my_package/__init__.py
my_package/__main__.py
my_package-0.0.1a6.dist-info/LICENSE.md
my_package-0.0.1a6.dist-info/METADATA
my_package-0.0.1a6.dist-info/WHEEL
my_package-0.0.1a6.dist-info/entry_points.txt
my_package-0.0.1a6.dist-info/top_level.txt
my_package-0.0.1a6.dist-info/RECORD
Note: you need to create your own check-whl.py script to produce the
above output. It is just three lines:
from zipfile import ZipFile
path = "dist/my_package-0.0.1a6-py3-none-any.whl" # <-- CHANGE
print('\n'.join(ZipFile(path).namelist()))
As expected, the bdist is missing the data files.
The _my_data folder is completely missing.
What if I create a _my_data/__init__.py? I repeat the
experiment and I find the data files are still not there! The
_my_data/ folder is included but it does not contain the data
files!
Solution
Contrary to the experience of others, this does work:
setuptools.setup(
...
packages=setuptools.find_packages(),
include_package_data=True, # <-- adds data files to bdist
...
)
With the fix in place, redo the experiment:
$ rm -fr build/ dist/ my_package.egg-info/
$ check-manifest
$ python.exe setup.py sdist bdist_wheel
Make sure the sdist still has the data files:
$ tar --list -f dist/my_package-0.0.1a6.tar.gz
my_package-0.0.1a6/
...
my_package-0.0.1a6/my_package/__init__.py
my_package-0.0.1a6/my_package/__main__.py
my_package-0.0.1a6/my_package/_my_data/
my_package-0.0.1a6/my_package/_my_data/consola.ttf <-- yay!
my_package-0.0.1a6/my_package/_my_data/icon.png <-- yay!
...
Look at the contents of the bdist:
$ python check-whl.py
my_package/__init__.py
my_package/__main__.py
my_package/_my_data/consola.ttf <--- yay!
my_package/_my_data/icon.png <--- yay!
my_package-0.0.1a6.dist-info/LICENSE.md
my_package-0.0.1a6.dist-info/METADATA
my_package-0.0.1a6.dist-info/WHEEL
my_package-0.0.1a6.dist-info/entry_points.txt
my_package-0.0.1a6.dist-info/top_level.txt
my_package-0.0.1a6.dist-info/RECORD
How not to test if data files are included
I recommend troubleshooting/testing using the approach outlined
above to inspect the sdist and bdist.
pip install in editable mode is not a valid test
Note: pip install -e . does not show if data files are
included in the bdist.
The symbolic link causes the installation to behave as if the
data files are included (because they already exist locally on
the developer's computer).
After pip install my_package, the data files are in the
virtual environment's lib/site-packages/my_package/ folder,
using the exact same file structure shown above in the list of
the whl contents.
Publishing to TestPyPI is a slow way to test
Publishing to TestPyPI and then installing and looking in
lib/site-packages/my_packages is a valid test, but it is too
time-consuming.
Ancient question and yet... package management of python really leaves a lot to be desired. So I had the use case of installing using pip locally to a specified directory and was surprised both package_data and data_files paths did not work out. I was not keen on adding yet another file to the repo so I ended up leveraging data_files and setup.py option --install-data; something like this
pip install . --install-option="--install-data=$PWD/package" -t package
Moving the folder containing the package data into to module folder solved the problem for me.
See this question: MANIFEST.in ignored on "python setup.py install" - no data files installed?
Just remove the line:
include_package_data=True,
from your setup script, and it will work fine. (Tested just now with latest setuptools.)
Like others in this thread, I'm more than a little surprised at the combination of longevity and still a lack of clarity, BUT the best answer for me was using check-manifest as recommended in the answer from #mike-gazes
So, using just a setup.cfg and no setup.py and additional text and python files required in the package, what worked for me was keeping this in setup.cfg:
[options]
packages = find:
include_package_data = true
and updating the MANIFEST.in based on the check-manifest output:
include *.in
include *.txt
include *.yml
include LICENSE
include tox.ini
recursive-include mypkg *.py
recursive-include mypkg *.txt
For a directory structure like:
foo/
├── foo
│ ├── __init__.py
│ ├── a.py
│ └── data.txt
└── setup.py
and setup.py
#!/usr/bin/env python
# -*- coding: utf-8 -*-
from setuptools import setup
NAME = 'foo'
DESCRIPTION = 'Test library to check how setuptools works'
URL = 'https://none.com'
EMAIL = 'gzorp#bzorp.com'
AUTHOR = 'KT'
REQUIRES_PYTHON = '>=3.6.0'
setup(
name=NAME,
version='0.0.0',
description=DESCRIPTION,
author=AUTHOR,
author_email=EMAIL,
python_requires=REQUIRES_PYTHON,
url=URL,
license='MIT',
classifiers=[
'Programming Language :: Python',
'Programming Language :: Python :: 3',
'Programming Language :: Python :: 3.6',
],
packages=['foo'],
package_data={'foo': ['data.txt']},
include_package_data=True,
install_requires=[],
extras_require={},
cmdclass={},
)
python setup.py bdist_wheel works.
Starting with Setuptools 62.3.0, you can now use recursive wildcards ("**") to include a (sub)directory recursively. This way you can include whole folders with all their folders and files in it.
For example, when using a pyproject.toml file, this is how you include two folders recursively:
[tool.setuptools.package-data]
"ema_workbench.examples.data" = ["**"]
"ema_workbench.examples.models" = ["**"]
But you can also only include certain file-types, in a folder and all subfolders. If you want to include all markdown (.md) files for example:
[tool.setuptools.package-data]
"ema_workbench.examples.data" = ["**/*.md"]
It should also work when using setup.py or setup.cfg.
See https://github.com/pypa/setuptools/pull/3309 for the details.
I have a problem when trying to install a Python package that I have created.
The package includes a bitmap images which is used within the package (for OCR).
My folder structure is the following:
mypackage
- mypackage
- media
- template.bmp
- module1.py
- module2.py
- etc...
- tests
- MANIFEST.in
- setup.py
template.bmp is used by the module1.py.
The MANIFEST.in file:
include mypackage/media/template.bmp
The setup.py:
setup(
....
packages = find_packages(exclude=["*.tests", "*.tests.*", "tests.*", "tests"]),
include_package_data=True,
package_data={'mypackage': ['media/template.bmp']},
...
)
When I run
python setup.py sdist
I can verify that the media folder is included along with template.bmp in the .egg file. However, when referencing the bitmap in a module using
directory = os.path.dirname(os.path.abspath(__file__))
template_path = directory + '/media/template.bmp'
cv2.imread(template_path, 0)
I get a file not found error. The directory variable is the following:
'C:\\anaconda3\\lib\\site-packages\\mypackage-0.0.1-py3.6.egg\\mypackage'
Am I missing something?
Using pkg_resources solved my problem.
template_path = pkg_resources.resource_filename(__name__, '/media/template.bmp')
So I have made a small application that I typically want to package and share it with the world. I read the tutorials for packaging my app using setuptools. It was going well until I got stuck at one place.
My directory structure is something like this:
- MyApp
- setup.py
- README.md
- LICENSE.txt
- MANIFEST.in
- myapp
- __init__.py
- index.py
- utils
- __init__.py
- helper.py
- some_dump.pickle
- images
- folder_1
- image_1.jpg
- image_2.jpg
- folder_2
- another_image_1.jpg
- another_image_2.jpg
- sounds
- sound_1.wav
- sound_2.wav
Given this directory structure, I have created the setup.py file with the packages option as the list containing myapp and myapp.utils. I have also created a MANIFEST.in file that will have all the files that I want to package but which are not included by default. Here is the content of the manifest file:
include README.md
include LICENSE.txt
include myapp/utils/*.pickle
include myapp/utils/sounds/*.wav
include myapp/utils/images/folder_1/*.jpg
include myapp/utils/images/folder_2/*.jpg
The problem is that all these files are created as required in the distribution package (.zip), but when I try to use the distribution package to try the install the package myself, I don't see the images, sounds or the pickle file.
i.e. when I run python setup.py install, I don't see these files in the build/lib folder. I am not sure what the reason is. Could anyone help me out with this?
Thanks
The manifest is about distribution, not about installation.
You need to specify in setup.py what to install and WHERE:
setup(...,
data_files=[(whereto1, [file1, file2]),
(whereto2, [file3])]
)
See http://docs.python.org/3/distutils/setupscript.html#installing-additional-files