The standard library clearly documents how to import source files directly (given the absolute file path to the source file), but this approach does not work if that source file uses implicit sibling imports as described in the example below.
How could that example be adapted to work in the presence of implicit sibling imports?
I already checked out this and this other Stackoverflow questions on the topic, but they do not address implicit sibling imports within the file being imported by hand.
Here's an illustrative example
Directory structure:
root/
- directory/
- app.py
- folder/
- implicit_sibling_import.py
- lib.py
app.py
:
import os
import importlib.util
# construct absolute paths
root = os.path.abspath(os.path.dirname(os.path.dirname(os.path.realpath(__file__))))
isi_path = os.path.join(root, 'folder', 'implicit_sibling_import.py')
def path_import(absolute_path):
'''implementation taken from https://docs.python.org/3/library/importlib.html#importing-a-source-file-directly'''
spec = importlib.util.spec_from_file_location(absolute_path, absolute_path)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
isi = path_import(isi_path)
print(isi.hello_wrapper())
lib.py
:
def hello():
return 'world'
implicit_sibling_import.py
:
import lib # this is the implicit sibling import. grabs root/folder/lib.py
def hello_wrapper():
return "ISI says: " + lib.hello()
#if __name__ == '__main__':
# print(hello_wrapper())
Running python folder/implicit_sibling_import.py
with the if __name__ == '__main__':
block commented out yields ISI says: world
in Python 3.6.
But running python directory/app.py
yields:
Traceback (most recent call last):
File "directory/app.py", line 10, in <module>
spec.loader.exec_module(module)
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 205, in _call_with_frames_removed
File "/Users/pedro/test/folder/implicit_sibling_import.py", line 1, in <module>
import lib
ModuleNotFoundError: No module named 'lib'
If I add import sys; sys.path.insert(0, os.path.dirname(isi_path))
to app.py
, python app.py
yields world
as intended, but I would like to avoid munging the sys.path
if possible.
I'd like python app.py
to print ISI says: world
and I'd like to accomplish this by modifying the path_import
function.
I'm not sure of the implications of mangling sys.path
. Eg. if there was directory/requests.py
and I added the path to directory
to the sys.path
, I wouldn't want import requests
to start importing directory/requests.py
instead of importing the requests library that I installed with pip install requests
.
The solution MUST be implemented as a python function that accepts the absolute file path to the desired module and returns the module object.
Ideally, the solution should not introduce side-effects (eg. if it does modify sys.path
, it should return sys.path
to its original state). If the solution does introduce side-effects, it should explain why a solution cannot be achieved without introducing side-effects.
PYTHONPATH
If I have multiple projects doing this, I don't want to have to remember to set PYTHONPATH
every time I switch between them. The user should just be able to pip install
my project and run it without any additional setup.
-m
The -m
flag is the recommended/pythonic approach, but the standard library also clearly documents How to import source files directly. I'd like to know how I can adapt that approach to cope with implicit relative imports. Clearly, Python's internals must do this, so how do the internals differ from the "import source files directly" documentation?
The easiest solution I could come up with is to temporarily modify sys.path
in the function doing the import:
from contextlib import contextmanager
@contextmanager
def add_to_path(p):
import sys
old_path = sys.path
sys.path = sys.path[:]
sys.path.insert(0, p)
try:
yield
finally:
sys.path = old_path
def path_import(absolute_path):
'''implementation taken from https://docs.python.org/3/library/importlib.html#importing-a-source-file-directly'''
with add_to_path(os.path.dirname(absolute_path)):
spec = importlib.util.spec_from_file_location(absolute_path, absolute_path)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
This should not cause any problems unless you do imports in another thread concurrently. Otherwise, since sys.path
is restored to its previous state, there should be no unwanted side effects.
Edit:
I realize that my answer is somewhat unsatisfactory but, digging into the code reveals that, the line spec.loader.exec_module(module)
basically results in exec(spec.loader.get_code(module.__name__),module.__dict__)
getting called. Here spec.loader.get_code(module.__name__)
is simply the code contained in lib.py.
Thus a better answer to the question would have to find a way to make the import
statement behave differently by simply injecting one or more global variables through the second argument of the exec-statement. However, "whatever you do to make the import machinery look in that file's folder, it'll have to linger beyond the duration of the initial import, since functions from that file might perform further imports when you call them", as stated by @user2357112 in the question comments.
Unfortunately the only way to change the behavior of the import
statement seems to be to change sys.path
or in a package __path__
. module.__dict__
already contains __path__
so that doesn't seem to work which leaves sys.path
(Or trying to figure out why exec does not treat the code as a package even though it has __path__
and __package__
... - But I don't know where to start - Maybe it has something to do with having no __init__.py
file).
Furthermore this issue does not seem to be specific to importlib
but rather a general problem with sibling imports.
Edit2: If you don't want the module to end up in sys.modules
the following should work (Note that any modules added to sys.modules
during the import are removed):
from contextlib import contextmanager
@contextmanager
def add_to_path(p):
import sys
old_path = sys.path
old_modules = sys.modules
sys.modules = old_modules.copy()
sys.path = sys.path[:]
sys.path.insert(0, p)
try:
yield
finally:
sys.path = old_path
sys.modules = old_modules