Skip to content

importnb imports notebooks as python modules.¤

if you're here, then there is a chance you have a notebook (.ipynb) in a directory saved as Untitled.ipynb. it is just sitting there, but what if it could be used as a python module? importnb is here to answer that question.

basic example¤

use importnb's Notebook finder and loader to import notebooks as modules

# with the new api
from importnb import imports
with imports("ipynb"):
    import Untitled

# with the explicit api
from importnb import imports
with Notebook():
    import Untitled

What does this snippet do?¤

the snippet begins with a context manager that modifies the files python can discover. it will find the Untitled.ipynb notebook and import it as a module with __name__ Untitled. the __file__ description will have .ipynb as an extension.

maybe when we give notebooks new life they eventually earn a better name than Untitled?

run a notebook as a script¤

the importnb command line interface mimics python's. it permits running notebooks files, modules, and raw json data.

the commands below execute a notebook module and file respectively.

importnb -m Untitled         # call the Untitled module as __main__
importnb Untitled.ipynb      # call the Untitled file as __main__

installing importnb¤

use either pip or conda/mamba

pip install importnb
conda install -cconda-forge importnb
mamba install -cconda-forge importnb

importnb features¤

  • importnb.Notebook offers parameters to customize how modules are imported
  • imports Jupyter notebooks as python modules
  • fuzzy finding conventions for finding files that are not valid python names
  • works with top-level await statements
  • integration with pytest
  • extensible machinery and entry points
  • translates Jupyter notebook files (ie .ipynb files) line-for-line to python source providing natural error messages
  • command line interface for running notebooks as python scripts
  • has no required dependencies

customizing parameters¤

the Notebook object has a few features that can be toggled:

  • lazy:bool=False lazy load the module, the namespace is populated when the module is access the first time.
  • position:int=0 the relative position of the import loader in the sys.path_hooks
  • fuzzy:bool=True use fuzzy searching syntax when underscores are encountered.
  • include_markdown_docstring:bool=True markdown blocks preceding function/class defs become docstrings.
  • include_magic:bool=True ignore any ipython magic syntaxes
  • only_defs:bool=False import only function and class definitions. ignore intermediate * expressions.
  • no_magic:bool=False execute IPython magic statements from the loader.

these features are defined in the importnb.loader.Interface class and they can be controlled throught the command line interface.

importing notebooks¤

the primary goal of this library is to make it easy to reuse python code in notebooks. below are a few ways to invoke python's import system within the context manager.

with importnb.imports("ipynb"):
    import Untitled
    import Untitled as nb
    __import__("Untitled")
    from importlib import import_module
    import_module("Untitled")

import data files¤

there is support for discovering data files. when discovered, data from disk on loaded and stored on the module with rich reprs.

with importnb.imports("toml", "json", "yaml"):
    pass

all the available entry points are found with

from importnb.entry_points import list_aliases
list_aliases()

loading directly from file¤

Untitled = Notebook.load("Untitled.ipynb")

fuzzy finding¤

often notebooks have names that are not valid python files names that are restricted alphanumeric characters and an _. the importnb fuzzy finder converts python's import convention into globs that will find modules matching specific patters. consider the statement:

with importnb.Notebook():
    import U_titl__d                        # U*titl**d.ipynb

importnb translates U_titl__d to a glob format that matches the pattern U*titl**d.ipynb when searching for the source. that means that importnb should fine Untitled.ipynb as the source for the import[^unless].

with importnb.Notebook():
    import _ntitled                        # *ntitled.ipynb
    import __d                     # **d.ipynb
    import U__                        # U**.ipynb

a primary motivation for this feature is name notebooks as if they were blog posts using the YYYY-MM-DD-title-here.ipynb convention. there are a few ways we could this file explicitly. the fuzzy finder syntax could like any of the following:

with importnb.Notebook():
    import __title_here
    import YYYY_MM_DD_title_here
    import __MM_DD_title_here

fuzzy name ambiguity¤

it is possible that a fuzzy import may be ambiguous are return multiple files. the importnb fuzzy finder will prefer the most recently changed file.

ambiguity can be avoided by using more explicit fuzzy imports that will reduce collisions. another option is use python's explicit import functions.

with importnb.Notebook():
    __import__("YYYY-MM-DD-title-here")
    import_module("YYYY-MM-DD-title-here")

importing your most recently changed notebook¤

an outcome of resolving the most recently changed is that you can import your most recent notebook with:

    import __                        # **.ipynb

integrations¤

pytest¤

since importnb transforms notebooks to python documents we can use these as source for tests. importnbs pytest extension is not fancy, it only allows for conventional pytest test discovery.

nbval is alternative testing tools that validates notebook outputs. this style is near to using notebooks as doctest while importnb primarily adds the ability to write unittests in notebooks. adding tests to notebooks help preserve them over time.

extensible¤

the importnb.Notebook machinery is extensible. it allows other file formats to be used. for example, pidgy uses importnb to import markdown files as compiled python code.

class MyLoader(importnb.Notebook): pass

developer¤

pip install -e.      # install in development mode
hatch run test:cov   # test 
  • importnb uses hatch for testing in python and IPython

appendix¤

line-for-line translation and natural error messages¤

a challenge with Jupyter notebooks is that they are json data. this poses problems:

  1. every valid line of code in a Jupyter notebook is a quoted json string
  2. json parsers don't have a reason to return line numbers.

the problem with quoted code¤

line-for-line json parser¤

python's json module is not pluggable in the way we need to find line numbers. since importnb is meant to be dependency free on installation we couldn't look to any other packages like ujson or json5.

the need for line numbers is enough that we ship a standalone json grammar parser. to do this without extra dependencies we use the lark grammar package at build time: * we've defined a json.gramar * we use hatch hooks to invoke lark-standalone that generates a standalone parser for the grammar. the generated file is shipped with the package.

the result of importnb is json data translated into vertically sparse, valid python code.

reproducibility caution with the fuzzy finder¤

⚠️ fuzzy finding is not reproducible as your system will change over time. in python, "explicit is better than implicit" so defining strong fuzzy strings is best practice if you MUST use esotric names. an alternative option is to use the importlib.import_module machinery