Using IPython notebooks under version control

问题:

What is a good strategy for keeping IPython notebooks under version control?

The notebook format is quite amenable for version control: if one wants to version control the notebook and the outputs then this works quite well. The annoyance comes when one wants only to version control the input, excluding the cell outputs (aka. "build products") which can be large binary blobs, especially for movies and plots. In particular, I am trying to find a good workflow that:

• allows me to choose between including or excluding output,
• prevents me from accidentally committing output if I do not want it,
• allows me to keep output in my local version,
• allows me to see when I have changes in the inputs using my version control system (i.e. if I only version control the inputs but my local file has outputs, then I would like to be able to see if the inputs have changed (requiring a commit). Using the version control status command will always register a difference since the local file has outputs.)
• allows me to update my working notebook (which contains the output) from an updated clean notebook. (update)

As mentioned, if I chose to include the outputs (which is desirable when using nbviewer for example), then everything is fine. The problem is when I do not want to version control the output. There are some tools and scripts for stripping the output of the notebook, but frequently I encounter the following issues:

1. I accidentally commit a version with the the output, thereby polluting my repository.
2. I clear output to use version control, but would really rather keep the output in my local copy (sometimes it takes a while to reproduce for example).
3. Some of the scripts that strip output change the format slightly compared to the Cell/All Output/Clear menu option, thereby creating unwanted noise in the diffs. This is resolved by some of the answers.
4. When pulling changes to a clean version of the file, I need to find some way of incorporating those changes in my working notebook without having to rerun everything. (update)

I have considered several options that I shall discuss below, but have yet to find a good comprehensive solution. A full solution might require some changes to IPython, or may rely on some simple external scripts. I currently use mercurial, but would like a solution that also works with git: an ideal solution would be version-control agnostic.

This issue has been discussed many times, but there is no definitive or clear solution from the user's perspective. The answer to this question should provide the definitive strategy. It is fine if it requires a recent (even development) version of IPython or an easily installed extension.

Update: I have been playing with my modified notebook version which optionally saves a .clean version with every save using Gregory Crosswhite's suggestions. This satisfies most of my constraints but leaves the following unresolved:

1. This is not yet a standard solution (requires a modification of the ipython source. Is there a way of achieving this behaviour with a simple extension? Needs some sort of on-save hook.
2. A problem I have with the current workflow is pulling changes. These will come in to the .clean file, and then need to be integrated somehow into my working version. (Of course, I can always re-execute the notebook, but this can be a pain, especially if some of the results depend on long calculations, parallel computations, etc.) I do not have a good idea about how to resolve this yet. Perhaps a workflow involving an extension like ipycache might work, but that seems a little too complicated.

Notes

Removing (stripping) Output

• When the notebook is running, one can use the Cell/All Output/Clear menu option for removing the output.
• There are some scripts for removing output, such as the script nbstripout.py which remove the output, but does not produce the same output as using the notebook interface. This was eventually included in the ipython/nbconvert repo, but this has been closed stating that the changes are now included in ipython/ipython,but the corresponding functionality seems not to have been included yet. (update) That being said, Gregory Crosswhite's solution shows that this is pretty easy to do, even without invoking ipython/nbconvert, so this approach is probably workable if it can be properly hooked in. (Attaching it to each version control system, however, does not seem like a good idea — this should somehow hook in to the notebook mechanism.)

Newsgroups

• Thoughts on the notebook format for version control.

Issues

• 977: Notebook feature requests (Open).
• 1280: Clear-all on save option (Open). (Follows from this discussion.)
• 3295: autoexported notebooks: only export explicitly marked cells (Closed). Resolved by extension 11 Add writeandexecute magic (Merged).

Pull Requests

• 1621: clear In[] prompt numbers on "Clear All Output" (Merged). (See also 2519 (Merged).)
• 1563: clear_output improvements (Merged).
• 3065: diff-ability of notebooks (Closed).
• 3291: Add the option to skip output cells when saving. (Closed). This seems extremely relevant, however was closed with the suggestion to use a "clean/smudge" filter. A relevant question what can you use if you want to strip off output before running git diff? seems not to have been answered.
• 3312: WIP: Notebook save hooks (Closed).
• 3747: ipynb -> ipynb transformer (Closed). This is rebased in 4175.
• 4175: nbconvert: Jinjaless exporter base (Merged).
• 142: Use STDIN in nbstripout if no input is given (Open).

回答1:

Here is my solution with git. It allows you to just add and commit (and diff) as usual: those operations will not alter your working tree, and at the same time (re)running a notebook will not alter your git history.

Although this can probably be adapted to other VCSs, I know it doesn't satisfy your requirements (at least the VSC agnosticity). Still, it is perfect for me, and although it's nothing particularly brilliant, and many people probably already use it, I didn't find clear instructions about how to implement it by googling around. So it may be useful to other people.

1. Save a file with this content somewhere (for the following, let us assume ~/bin/ipynb_output_filter.py)
2. Make it executable (chmod +x ~/bin/ipynb_output_filter.py)

3. Create the file ~/.gitattributes, with the following content

   *.ipynb    filter=dropoutput_ipynb
   

4. Run the following commands:

   git config --global core.attributesfile ~/.gitattributes
   git config --global filter.dropoutput_ipynb.clean ~/bin/ipynb_output_filter.py
   git config --global filter.dropoutput_ipynb.smudge cat
   

Done!

Limitations:

• it works only with git
• in git, if you are in branch somebranch and you do git checkout otherbranch; git checkout somebranch, you usually expect the working tree to be unchanged. Here instead you will have lost the output and cells numbering of notebooks whose source differs between the two branches.
• more in general, the output is not versioned at all, as with Gregory's solution. In order to not just throw it away every time you do anything involving a checkout, the approach could be changed by storing it in separate files (but notice that at the time the above code is run, the commit id is not known!), and possibly versioning them (but notice this would require something more than a git commit notebook_file.ipynb, although it would at least keep git diff notebook_file.ipynb free from base64 garbage).
• that said, incidentally if you do pull code (i.e. committed by someone else not using this approach) which contains some output, the output is checked out normally. Only the locally produced output is lost.

My solution reflects the fact that I personally don't like to keep generated stuff versioned - notice that doing merges involving the output is almost guaranteed to invalidate the output or your productivity or both.

EDIT:

• if you do adopt the solution as I suggested it - that is, globally - you will have trouble in case for some git repo you want to version output. So if you want to disable the output filtering for a specific git repository, simply create inside it a file .git/info/attributes, with

  **.ipynb filter=

as content. Clearly, in the same way it is possible to do the opposite: enable the filtering only for a specific repository.

• the code is now maintained in its own git repo

• if the instructions above result in ImportErrors, try adding "ipython" before the path of the script:

  git config --global filter.dropoutput_ipynb.clean ipython ~/bin/ipynb_output_filter.py
  

EDIT: May 2016 (updated February 2017): there are several alternatives to my script - for completeness, here is a list of those I know: nbstripout (other variants), nbstrip, jq.

回答2:

We have a collaborative project where the product is Jupyter Notebooks, and we've use an approach for the last six months that is working great: we activate saving the .py files automatically and track both .ipynb files and the .py files.

That way if someone wants to view/download the latest notebook they can do that via github or nbviewer, and if someone wants to see how the the notebook code has changed, they can just look at the changes to the .py files.

For Jupyter notebook servers, this can be accomplished by adding the lines

import os
from subprocess import check_call

def post_save(model, os_path, contents_manager):
    """post-save hook for converting notebooks to .py scripts"""
    if model['type'] != 'notebook':
        return # only do this for notebooks
    d, fname = os.path.split(os_path)
    check_call(['jupyter', 'nbconvert', '--to', 'script', fname], cwd=d)

c.FileContentsManager.post_save_hook = post_save

to the jupyter_notebook_config.py file and restarting the notebook server.

If you aren't sure in which directory to find your jupyter_notebook_config.py file, you can type jupyter --config-dir, and if you don't find the file there, you can create it by typing jupyter notebook --generate-config.

For Ipython 3 notebook servers, this can be accomplished by adding the lines

import os
from subprocess import check_call

def post_save(model, os_path, contents_manager):
    """post-save hook for converting notebooks to .py scripts"""
    if model['type'] != 'notebook':
        return # only do this for notebooks
    d, fname = os.path.split(os_path)
    check_call(['ipython', 'nbconvert', '--to', 'script', fname], cwd=d)

c.FileContentsManager.post_save_hook = post_save

to the ipython_notebook_config.py file and restarting the notebook server. These lines are from a github issues answer @minrk provided and @dror includes them in his SO answer as well.

For Ipython 2 notebook servers, this can be accomplished by starting the server using:

ipython notebook --script

or by adding the line

c.FileNotebookManager.save_script = True

to the ipython_notebook_config.py file and restarting the notebook server.

If you aren't sure in which directory to find your ipython_notebook_config.py file, you can type ipython locate profile default, and if you don't find the file there, you can create it by typing ipython profile create.

Here's our project on github that is using this approach: and here's a github example of exploring recent changes to a notebook.

We've been very happy with this.

回答3:

I have created nbstripout, based on MinRKs gist, which supports both Git and Mercurial (thanks to mforbes). It is intended to be used either standalone on the command line or as a filter, which is easily (un)installed in the current repository via nbstripout install / nbstripout uninstall.

Get it from PyPI or simply

pip install nbstripout

回答4:

Here is a new solution from Cyrille Rossant for IPython 3.0, which persists to markdown files rather than json-based ipymd files:

https://github.com/rossant/ipymd

回答5:

(2017-02)

strategies

• on_commit():
  • strip the output > name.ipynb (nbstripout, )
  • strip the output > name.clean.ipynb (nbstripout,)
  • always nbconvert to python: name.ipynb.py (nbconvert)
  • always convert to markdown: name.ipynb.md (nbconvert, ipymd)
• vcs.configure():
  • git difftool, mergetool: nbdiff and nbmerge from nbdime

tools

• nbstripout: strip the outputs from a notebook
  • src: https://gist.github.com/minrk/6176788
  • src: https://github.com/kynan/nbstripout
    • pip install nbstripout; nbstripout install
• ipynb_output_filter: strip the outputs from a notebook
  • src: https://github.com/toobaz/ipynb_output_filter/blob/master/ipynb_output_filter.py
• ipymd: convert between {Jupyter, Markdown, O'Reilly Atlas Markdown, OpenDocument, .py}
  • src: https://github.com/rossant/ipymd
• nbdime: "Tools for diffing and merging of Jupyter notebooks." (2015)
  • src: https://github.com/jupyter/nbdime
  • docs: http://nbdime.readthedocs.io/
    • nbdiff: compare notebooks in a terminal-friendly way
      • nbdime nbdiff works as a git diff tool: https://nbdime.readthedocs.io/en/latest/#git-integration-quickstart
    • nbmerge: three-way merge of notebooks with automatic conflict resolution
      • nbdime nbmerge works as a git merge tool
    • nbdiff-web: shows you a rich rendered diff of notebooks
    • nbmerge-web: gives you a web-based three-way merge tool for notebooks
    • nbshow: present a single notebook in a terminal-friendly way

回答6:

As pointed out by, the --script is deprecated in 3.x. This approach can be used by applying a post-save-hook. In particular, add the following to ipython_notebook_config.py:

import os
from subprocess import check_call

def post_save(model, os_path, contents_manager):
    """post-save hook for converting notebooks to .py scripts"""
    if model['type'] != 'notebook':
        return # only do this for notebooks
    d, fname = os.path.split(os_path)
    check_call(['ipython', 'nbconvert', '--to', 'script', fname], cwd=d)

c.FileContentsManager.post_save_hook = post_save

The code is taken from #8009.

回答7:

I use a very pragmatic approach; which work well for several notebooks, at several sides. And it even enables me to 'transfer' notebooks around. It works both for Windows as Unix/MacOS.
Al thought it is simple, is solve the problems above...

Concept

Basically, do not track the .ipnyb-files, only the corresponding .py-files.
By starting the notebook-server with the --script option, that file is automatically created/saved when the notebook is saved.

Those .py-files do contain all input; non-code is saved into comments, as are the cell-borders. Those file can be read/imported ( and dragged) into the notebook-server to (re)create a notebook. Only the output is gone; until it is re-run.

Personally I use mercurial to version-track the .py files; and use the normal (command-line) commands to add, check-in (ect) for that. Most other (D)VCS will allow this to.

Its simple to track the history now; the .py are small, textual and simple to diff. Once and a while, we need a clone (just branch; start a 2nd notebook-sever there), or a older version (check-it out and import into a notebook-server), etc.

Tips & tricks

• Add *.ipynb to '.hgignore', so Mercurial knows it can ignore those files
• Create a (bash) script to start the server (with the --script option) and do version-track it
• Saving a notebook does save the .py-file, but does not check it in.
  • This is a drawback: One can forget that
  • It's a feature also: It possible to save a notebook (and continue later) without clustering the repository-history.

Wishes

• It would be nice to have a buttons for check-in/add/etc in the notebook Dashboard
• A checkout to (by example) file@date+rev.py) should be helpful It would be to much work to add that; and maybe I will do so once. Until now, I just do that by hand.

回答8:

Unfortunately, I do not know much about Mercurial, but I can give you a possible solution that works with Git, in the hopes that you might be able to translate my Git commands into their Mercurial equivalents.

For background, in Git the add command stores the changes that have been made to a file into a staging area. Once you have done this, any subsequent changes to the file are ignored by Git unless you tell it to stage them as well. Hence, the following script, which, for each of the given files, strips out all of the outputs and prompt_number sections, stages the stripped file, and then restores the original:

NOTE: If running this gets you an error message like ImportError: No module named IPython.nbformat, then use ipython to run the script instead of python.

from IPython.nbformat import current
import io
from os import remove, rename
from shutil import copyfile
from subprocess import Popen
from sys import argv

for filename in argv[1:]:
    # Backup the current file
    backup_filename = filename + ".backup"
    copyfile(filename,backup_filename)

    try:
        # Read in the notebook
        with io.open(filename,'r',encoding='utf-8') as f:
            notebook = current.reads(f.read(),format="ipynb")

        # Strip out all of the output and prompt_number sections
        for worksheet in notebook["worksheets"]:
            for cell in worksheet["cells"]:
               cell.outputs = []
               if "prompt_number" in cell:
                    del cell["prompt_number"]

        # Write the stripped file
        with io.open(filename, 'w', encoding='utf-8') as f:
            current.write(notebook,f,format='ipynb')

        # Run git add to stage the non-output changes
        print("git add",filename)
        Popen(["git","add",filename]).wait()

    finally:
        # Restore the original file;  remove is needed in case
        # we are running in windows.
        remove(filename)
        rename(backup_filename,filename)

Once the script has been run on the files whose changes you wanted to commit, just run git commit.

回答9:

To follow up on the excellent script by Pietro Battiston, if you get a Unicode parsing error like this:

Traceback (most recent call last):
  File "/Users/kwisatz/bin/ipynb_output_filter.py", line 33, in <module>
write(json_in, sys.stdout, NO_CONVERT)
  File "/Users/kwisatz/anaconda/lib/python2.7/site-packages/IPython/nbformat/__init__.py", line 161, in write
fp.write(s)
UnicodeEncodeError: 'ascii' codec can't encode character u'\u2014' in position 11549: ordinal not in range(128)

You can add at the beginning of the script:

reload(sys)
sys.setdefaultencoding('utf8')

回答10:

After digging around, I finally found this relatively simple pre-save hook on the Jupyter docs. It strips the cell output data. You have to paste it into the jupyter_notebook_config.py file (see below for instructions).

def scrub_output_pre_save(model, **kwargs):
    """scrub output before saving notebooks"""
    # only run on notebooks
    if model['type'] != 'notebook':
        return
    # only run on nbformat v4
    if model['content']['nbformat'] != 4:
        return

    for cell in model['content']['cells']:
        if cell['cell_type'] != 'code':
            continue
        cell['outputs'] = []
        cell['execution_count'] = None
        # Added by binaryfunt:
        if 'collapsed' in cell['metadata']:
            cell['metadata'].pop('collapsed', 0)

c.FileContentsManager.pre_save_hook = scrub_output_pre_save

From Rich Signell's answer:

If you aren't sure in which directory to find your jupyter_notebook_config.py file, you can type jupyter --config-dir [into command prompt/terminal], and if you don't find the file there, you can create it by typing jupyter notebook --generate-config.

回答11:

I did what Albert & Rich did - Don't version .ipynb files (as these can contain images, which gets messy). Instead, either always run ipython notebook --script or put c.FileNotebookManager.save_script = True in your config file, so that a (versionable) .py file is always created when you save your notebook.

To regenerate notebooks (after checking out a repo or switching a branch) I put the script py_file_to_notebooks.py in the directory where I store my notebooks.

Now, after checking out a repo, just run python py_file_to_notebooks.py to generate the ipynb files. After switching branch, you may have to run python py_file_to_notebooks.py -ov to overwrite the existing ipynb files.

Just to be on the safe side, it's good to also add *.ipynb to your .gitignore file.

Edit: I no longer do this because (A) you have to regenerate your notebooks from py files every time you checkout a branch and (B) there's other stuff like markdown in notebooks that you lose. I instead strip output from notebooks using a git filter. Discussion on how to do this is here.

回答12:

Ok, so it looks like the current best solution, as per a discussion here, is to make a git filter to automatically strip output from ipynb files on commit.

Here's what I did to get it working (copied from that discussion):

I modified cfriedline's nbstripout file slightly to give an informative error when you can't import the latest IPython: https://github.com/petered/plato/blob/fb2f4e252f50c79768920d0e47b870a8d799e92b/notebooks/config/strip_notebook_output And added it to my repo, lets say in ./relative/path/to/strip_notebook_output

Also added the file .gitattributes file to the root of the repo, containing:

*.ipynb filter=stripoutput

And created a setup_git_filters.sh containing

git config filter.stripoutput.clean "$(git rev-parse --show-toplevel)/relative/path/to/strip_notebook_output" 
git config filter.stripoutput.smudge cat
git config filter.stripoutput.required true

And ran source setup_git_filters.sh. The fancy $(git rev-parse...) thing is to find the local path of your repo on any (Unix) machine.

回答13:

I've built python package that solves this problem

https://github.com/brookisme/gitnb

It provides a CLI with a git-inspired syntax to track/update/diff notebooks inside your git repo.

Heres' an example

# add a notebook to be tracked
gitnb add SomeNotebook.ipynb

# check the changes before commiting
gitnb diff SomeNotebook.ipynb

# commit your changes (to your git repo)
gitnb commit -am "I fixed a bug"

Note that last step, where I'm using "gitnb commit" is committing to your git repo. Its essentially a wrapper for

# get the latest changes from your python notebooks
gitnb update

# commit your changes ** this time with the native git commit **
git commit -am "I fixed a bug"

There are several more methods, and can be configured so that it requires more or less user input at each stage, but thats the general idea.

回答14:

I finally found a productive and simple way to make Jupyter and Git play nicely together. I'm still in the first steps, but I already think it is a lot better than all other convoluted solutions.

Visual Studio Code is a cool and open source code editor from Microsoft. It has an excellent Python extension that now allows you to import a Jupyter Notebook as python code.

After you import your notebook to a python file, all the code and markdown will be together in a ordinary python file, with special markers in comments. You can see in the image below:

Your python file just has the contents of the notebook input cells. The output will be generated in a split window. You have pure code in the notebook, it doesn't change while you just execute it. No mingled output with your code. No strange Json incomprehensible format to analyze your diffs.

Just pure python code where you can easily identify every single diff.

I don't even need to version my .ipynb files anymore. I can put a *.ipynb line in .gitignore.

Need to generate a notebook to publish or share with someone? No problem, just click the export button in the interactive python window

I've been using it just for a day, but finally I can happily use Jupyter with Git.

P.S.: VSCode code completion is a lot better than Jupyter.

回答15:

Just come across "jupytext" which looks like a perfect solution. It generates a .py file from the notebook and then keeps both in sync. You can version control, diff and merge inputs via the .py file without losing the outputs. When you open the notebook it uses the .py for input cells and the .ipynb for output. And if you want to include the output in git then you can just add the ipynb.

https://github.com/mwouts/jupytext

回答16:

How about the idea discussed in the post below, where the output of the notebook should be kept, with the argument that it might take a long time to generate it, and it is handy since GitHub can now render notebooks. There are auto-save hooks added for exporting .py file, used for diffs and .html for sharing with team members who do not use notebooks or git.

https://towardsdatascience.com/version-control-for-jupyter-notebook-3e6cef13392d

回答17:

This jupyter extension enables users to push jupyter notebooks directly to github.

Please look here

https://github.com/sat28/githubcommit