[PYTHON] Create a wheel of your own OpenCV module

I'm grafi working part-time at ALBERT Co., Ltd. I haven't posted to Qiita before, but I'll post it for the first time.

Motivation

When doing data analysis, I often read images with Python and perform simple processing, so I would like OpenCV. Most people're done with Anaconda, but I personally don't want to leave the Python ecosystem, and I'd love to specify which libraries to link with.

That's why you're building OpenCV. I think there are explanations about how to build it here and there, so I will omit it. This will create a file such as cv2. Somehow .so as a Python module. To install it globally, just do make install.

However, from the environment created with virtualenv (or venv), this cv2 module cannot be seen unless --system-site-packages is added. If I somehow copy cv2.so to site-packages in the virtualenv environment, it works, but it's just not cool.

So I wondered if I could make this a wheel.

How to do

What I really want to do is "I want to make one pre-built extension module into a wheel", but it seems that this is not supported.

Therefore, I made a lot of mistakes, but I managed to do it. I mainly referred to

• How can I make a Python Wheel from an existing native library? - Stack Overflow
• Package only binary compiled .so files of a python library compiled with Cython - Stack Overflow
• Including Data Files — setuptools 35.0.2 documentation

There are three.

First, the directory structure is as follows.

--+-- cv2 --+-- _native --+-- cv2.somehow.so
  |         |             |
  |         |             +-- __init__.py
  |         | 
  |         +-- __init__.py
  |
  +-- setup.py
  |
  +-- MANIFEST.in

In __init__.py directly under the cv2 directory

import * from _native.cv2

Write. The __init__.py under the _native directory can be empty.

setup.py is as follows

from setuptools import setup
from setuptools.dist import Distribution

class BinaryDistribution(Distribution):
    def has_ext_modules(foo):
        return True

setup(
    name='cv2',
    version='1.0',
    packages=['cv2'],
    include_package_data=True,
    distclass=BinaryDistribution,
)

Also, MANIFEST.in is

include cv2/_native/cv2.somehow.so

Write.

Now when you run python setup.py bdist_wheel, you will have a wheel with a file name like cv2-1.0-cp36-cp36m-linux_x86_64.whl under dist. I'm happy.

Commentary

It's simple to do, just create a dummy package and export all the definitions in the original cv2 module. I'd like to hide the original cv2 module for the time being, but if I renamecv2. Somehow .so to _cv2. Somehow .so, it will not work, so I solved it by putting it in another package. did.

By forcibly telling you that the extension module is included by setting has_ext_modules = True, you can make the wheel so that the installation will be rejected if you bring it to another OS.

application

numpy and scipy will not perform well unless you compile them yourself and link them to a high-performance matrix operation library, but it is also convenient to have the compiled result as a wheel. I will write this detail later.