First question, when is numpy needed? It is needed during the setup (i.e. when build_ext-funcionality is called) and in the installation, when the module is used. That means numpy should be in setup_requires and in install_requires.
There are following alternatives to solve the issue for the setup:
- using PEP 517/518 (which is more straight forward IMO)
- using
setup_requires-argument of setup and postponing import of numpy until setup's requirements are satisfied (which is not the case at the start of setup.py's execution)
PEP 517/518-solution:
Put next to setup.py a pyproject.toml-file , with the following content:
[build-system]
requires = ["setuptools", "wheel", "Cython>=0.29", "numpy >= 1.15"]
which defines packages needed for building, and then install using pip install . in the folder with setup.py. A disadvantage of this method is that python setup.py install no longer works, as it is pip that reads pyproject.toml. However, I would use this approach whenever possible.
Posponing import
This approach is more complicated and somewhat hacky, but works also without pip.
First, let's take a look at unsuccessful tries so far:
pybind11-trick
@chrisb's "pybind11"-trick, which can be found here: With help of an indirection, one delays the call to import numpy until numpy is present during the setup-phase, i.e.:
class get_numpy_include(object):
def __str__(self):
import numpy
return numpy.get_include()
...
my_c_lib_ext = setuptools.Extension(
...
include_dirs=[get_numpy_include()]
)
Clever! The problem: it doesn't work with the Cython-compiler: somewhere down the line, Cython passes the get_numpy_include-object to os.path.join(...,...) which checks whether the argument is really a string, which it obviously isn't.
This could be fixed by inheriting from str, but the above shows the dangers of the approach in the long run - it doesn't use the designed mechanics, is brittle and may easily fail in the future.
the classical build_ext-solution
Which looks as following:
...
from setuptools.command.build_ext import build_ext as _build_ext
class build_ext(_build_ext):
def finalize_options(self):
_build_ext.finalize_options(self)
# Prevent numpy from thinking it is still in its setup process:
__builtins__.__NUMPY_SETUP__ = False
import numpy
self.include_dirs.append(numpy.get_include())
setupttools.setup(
...
cmdclass={'build_ext':build_ext},
...
)
Yet also this solution doesn't work with cython-extensions, because pyx-files don't get recognized.
The real question is, how did pyx-files get recognized in the first place? The answer is this part of setuptools.command.build_ext:
...
try:
# Attempt to use Cython for building extensions, if available
from Cython.Distutils.build_ext import build_ext as _build_ext
# Additionally, assert that the compiler module will load
# also. Ref #1229.
__import__('Cython.Compiler.Main')
except ImportError:
_build_ext = _du_build_ext
...
That means setuptools tries to use the Cython's build_ext if possible, and because the import of the module is delayed until build_ext is called, it founds Cython present.
The situation is different when setuptools.command.build_ext is imported at the beginning of the setup.py - the Cython isn't yet present and a fall back without cython-functionality is used.
mixing up pybind11-trick and classical solution
So let's add an indirection, so we don't have to import setuptools.command.build_ext directly at the beginning of setup.py:
....
# factory function
def my_build_ext(pars):
# import delayed:
from setuptools.command.build_ext import build_ext as _build_ext#
# include_dirs adjusted:
class build_ext(_build_ext):
def finalize_options(self):
_build_ext.finalize_options(self)
# Prevent numpy from thinking it is still in its setup process:
__builtins__.__NUMPY_SETUP__ = False
import numpy
self.include_dirs.append(numpy.get_include())
#object returned:
return build_ext(pars)
...
setuptools.setup(
...
cmdclass={'build_ext' : my_build_ext},
...
)
