Okay, the scenario is very simple. I have this file structure:
.
├── interface.py
├── pkg
│ ├── __init__.py
│ ├── mod1.py
│ ├── mod2.py
Now, these are my conditions:
- mod2 needs to import mod1.
- both interface.py and mod2 needs to be run independently as a main script. If you want, think interface as the actual program and mod2 as an internal tester of the package.
So, in Python 2 I would simply do import mod1
inside mod2.py and both python2 mod2.py
and python2 interface.py
would work as expected.
However, and this is the part I less understand, using Python 3.5.2, if I do import mod1
; then I can do python3 mod2.py
, but python3 interface.py
throws: ImportError: No module named 'mod1'
:(
So, apparently, python 3 proposes to use import pkg.mod1
to avoid collisions against built-in modules. Ok, If I use that I can do python3 interface.py
; but then I can't python3 mod2.py
because: ImportError: No module named 'pkg'
Similarly, If I use relative import:
from . import mod1
then python3 interface.py
works; but mod2.py says SystemError: Parent module '' not loaded, cannot perform relative import
:( :(
The only "solution", I've found is to go up one folder and do python -m pkg.mod2
and then it works. But do we have to be adding the package prefix pkg
to every import to other modules within that package? Even more, to run any scripts inside the package, do I have to remember to go one folder up and use the -m switch? That's the only way to go??
I'm confused. This scenario was pretty straightforward with python 2, but looks awkward in python 3.
UPDATE: I have upload those files with the (referred as "solution" above) working source code here: https://gitlab.com/Akronix/test_python3_packages. Note that I still don't like it, and looks much uglier than the python2 solution.
Related SO questions I've already read:
- Python -- import the package in a module that is inside the same package
- How to do relative imports in Python?
- Absolute import module in same package
Related links:
TLDR:
python -m pkg.mod2
.from . import mod1
.Using the
-m
switch is indeed the "only" solution - it was already the only solution before. The old behaviour simply only ever worked out of sheer luck; it could be broken without even modifying your code.Going "one folder up" merely adds your package to the search path. Installing your package or modifying the search path works as well. See below for details.
You must have a reference to your package - otherwise it is ambiguous which module you want. The package reference can be either absolute or relative.
A relative import is usually what you want. It saves writing
pkg
explicitly, making it easier to refactor and move modules.Note that you can always use
<import> as <name>
to bind your import to a different name. For example,import pkg.mod2 as mod2
lets you work with just the module name.If your package is properly installed, you can use the
-m
switch from anywhere. For example, you can always usepython3 -m json.tool
.If your package is not installed (yet), you can set
PYTHONPATH
to its base directory. This includes your package in the search path, and allows the-m
switch to find it properly.If you are in the executable's directory, you can execute
export PYTHONPATH="$(pwd)/.."
to quickly mount the package for import.This scenario was basically broken in python 2. While it was straightforward in many cases, it was difficult or outright impossible to fix in any other cases.
The new behaviour is more awkward in the straightforward case, but robust and reliable in any case.