[python] Import local function from a module housed in another directory with relative imports in Jupyter Notebook using Python 3

I have a directory structure similar to the following

meta_project
    project1
        __init__.py
        lib
            module.py
            __init__.py
    notebook_folder
        notebook.jpynb

When working in notebook.jpynb if I try to use a relative import to access a function function() in module.py with:

from ..project1.lib.module import function

I get the following error:

SystemError                               Traceback (most recent call last)
<ipython-input-7-6393744d93ab> in <module>()
----> 1 from ..project1.lib.module import function

SystemError: Parent module '' not loaded, cannot perform relative import

Is there any way to get this to work using relative imports?

Note, the notebook server is instantiated at the level of the meta_project directory, so it should have access to the information in those files.

Note, also, that at least as originally intended project1 wasn't thought of as a module and therefore does not have an __init__.py file, it was just meant as a file-system directory. If the solution to the problem requires treating it as a module and including an __init__.py file (even a blank one) that is fine, but doing so is not enough to solve the problem.

I share this directory between machines and relative imports allow me to use the same code everywhere, & I often use notebooks for quick prototyping, so suggestions that involve hacking together absolute paths are unlikely to be helpful.

Edit: This is unlike Relative imports in Python 3, which talks about relative imports in Python 3 in general and – in particular – running a script from within a package directory. This has to do with working within a jupyter notebook trying to call a function in a local module in another directory which has both different general and particular aspects.

Here's my 2 cents:

import sys

map the path where the module file is located. In my case it was the desktop

sys.path.append('/Users/John/Desktop')

Either import the whole mapping module BUT then you have to use the .notation to map the classes like mapping.Shipping()

import mapping #mapping.py is the name of my module file

shipit = mapping.Shipment() #Shipment is the name of the class I need to use in the mapping module

Or import the specific class from the mapping module

from mapping import Mapping

shipit = Shipment() #Now you don't have to use the .notation

I have found that python-dotenv helps solve this issue pretty effectively. Your project structure ends up changing slightly, but the code in your notebook is a bit simpler and consistent across notebooks.

For your project, do a little install.

pipenv install python-dotenv

Then, project changes to:

+-- .env (this can be empty)
+-- ipynb
¦   +-- 20170609-Examine_Database_Requirements.ipynb
¦   +-- 20170609-Initial_Database_Connection.ipynb
+-- lib
    +-- __init__.py
    +-- postgres.py

And finally, your import changes to:

import os
import sys

from dotenv import find_dotenv


sys.path.append(os.path.dirname(find_dotenv()))

A +1 for this package is that your notebooks can be several directories deep. python-dotenv will find the closest one in a parent directory and use it. A +2 for this approach is that jupyter will load environment variables from the .env file on startup. Double whammy.

Researching this topic myself and having read the answers I recommend using the path.py library since it provides a context manager for changing the current working directory.

You then have something like

import path
if path.Path('../lib').isdir():
    with path.Path('..'):
        import lib

Although, you might just omit the isdir statement.

Here I'll add print statements to make it easy to follow what's happening

import path
import pandas

print(path.Path.getcwd())
print(path.Path('../lib').isdir())
if path.Path('../lib').isdir():
    with path.Path('..'):
        print(path.Path.getcwd())
        import lib
        print('Success!')
print(path.Path.getcwd())

which outputs in this example (where lib is at /home/jovyan/shared/notebooks/by-team/data-vis/demos/lib):

/home/jovyan/shared/notebooks/by-team/data-vis/demos/custom-chart
/home/jovyan/shared/notebooks/by-team/data-vis/demos
/home/jovyan/shared/notebooks/by-team/data-vis/demos/custom-chart

Since the solution uses a context manager, you are guaranteed to go back to your previous working directory, no matter what state your kernel was in before the cell and no matter what exceptions are thrown by importing your library code.

So far, the accepted answer has worked best for me. However, my concern has always been that there is a likely scenario where I might refactor the notebooks directory into subdirectories, requiring to change the module_path in every notebook. I decided to add a python file within each notebook directory to import the required modules.

Thus, having the following project structure:

project
|__notebooks
   |__explore
      |__ notebook1.ipynb
      |__ notebook2.ipynb
      |__ project_path.py
   |__ explain
       |__notebook1.ipynb
       |__project_path.py
|__lib
   |__ __init__.py
   |__ module.py

I added the file project_path.py in each notebook subdirectory (notebooks/explore and notebooks/explain). This file contains the code for relative imports (from @metakermit):

import sys
import os

module_path = os.path.abspath(os.path.join(os.pardir, os.pardir))
if module_path not in sys.path:
    sys.path.append(module_path)

This way, I just need to do relative imports within the project_path.py file, and not in the notebooks. The notebooks files would then just need to import project_path before importing lib. For example in 0.0-notebook.ipynb:

import project_path
import lib

The caveat here is that reversing the imports would not work. THIS DOES NOT WORK:

import lib
import project_path

Thus care must be taken during imports.

All other answers here depends on adding code the the notebook(!)

In my opinion is bad practice to hardcode a specific path into the notebook code, or otherwise depend on the location, since this makes it really hard to refactor you code later on. Instead I would recommend you to add the root project folder to PYTHONPATH when starting up your Jupyter notebook server, either directly from the project folder like so

env PYTHONPATH=`pwd` jupyter notebook

or if you are starting it up from somewhere else, use the absolute path like so

env PYTHONPATH=/Users/foo/bar/project/ jupyter notebook

I have just found this pretty solution:

import sys; sys.path.insert(0, '..') # add parent folder path where lib folder is
import lib.store_load # store_load is a file on my library folder

You just want some functions of that file

from lib.store_load import your_function_name

If python version >= 3.3 you do not need init.py file in the folder

Came here searching for best practices in abstracting code to submodules when working in Notebooks. I'm not sure that there is a best practice. I have been proposing this.

A project hierarchy as such:

+-- ipynb
¦   +-- 20170609-Examine_Database_Requirements.ipynb
¦   +-- 20170609-Initial_Database_Connection.ipynb
+-- lib
    +-- __init__.py
    +-- postgres.py

And from 20170609-Initial_Database_Connection.ipynb:

    In [1]: cd ..

    In [2]: from lib.postgres import database_connection

This works because by default the Jupyter Notebook can parse the cd command. Note that this does not make use of Python Notebook magic. It simply works without prepending %bash.

Considering that 99 times out of a 100 I am working in Docker using one of the Project Jupyter Docker images, the following modification is idempotent

    In [1]: cd /home/jovyan

    In [2]: from lib.postgres import database_connection