Documentation modules overview

How to contribute to jupyter notebooks

The documentation is built from notebooks in docs_src/. Follow the steps below to build documentation. For more information about generating and authoring notebooks, see fastai.gen_doc.gen_notebooks.


Generate and update notebook skeletons automatically from modules. Includes an overview of the whole authoring process.

Create HTML (jekyll) docs from notebooks.

Underlying documentation functions; most important is show_doc.

Process for contributing to the docs

If you want to help us and contribute to the docs, you just have to make modifications to the source notebooks, our scripts will then automatically convert them to HTML. There is just one script to run after cloning the fastai repo, to ensure that everything works properly. The rest of this page goes more in depth about all the functionalities this module offers, but if you just want to add a sentence or correct a typo, make a PR with the notebook changed and we'll take care of the rest.

Thing to run after git clone

Make sure you follow this recipe:

git clone
cd fastai

This will take care of everything that is explained in the following two sections. We'll tell you what they do, but you need to execute just this one script.

Note: windows users, not using bash emulation, will need to invoke the command as:

python tools\run-after-git-clone

If you're on windows, you also need to convert the Unix symlink between docs_src\imgs and docs\imgs. You will need to (1) remove docs_src\imgs, (2) execute cmd.exe as administrator, and (3) finally, in the docs_src folder, execute:

mklink /d imgs ..\docs\imgs

after-git-clone #1: a mandatory notebook strip out

Currently we only store source code cells under git (and a few extra fields for documentation notebooks). If you would like to commit or submit a PR, you need to confirm to that standard.

This is done automatically during diff/commit git operations, but you need to configure your local repository once to activate that instrumentation.

Therefore, your developing process will always start with:


The last command tells git to invoke configuration stored in fastai/.gitconfig, so your git diff and git commit invocations for this particular repository will now go via tools/fastai-nbstripout which will do all the work for you.

You don't need to run it if you run:


If you skip this configuration your commit/PR involving notebooks will not be accepted, since it'll carry in it many JSON bits which we don't want in the git repository. Those unwanted bits create collisions and lead to unnecessarily complicated and time wasting merge activities. So please do not skip this step.

Note: we can't make this happen automatically, since git will ignore a repository-stored .gitconfig for security reasons, unless a user will tell git to use it (and thus trust it).

If you'd like to check whether you already trusted git with using fastai/.gitconfig please look inside fastai/.git/config, which should have this entry:

        path = ../.gitconfig

or alternatively run:

tools/trust-origin-git-config -t

after-git-clone #2: automatically updating doc notebooks to be trusted on git pull

We want the doc notebooks to be already trusted when you load them in jupyter notebook, so this script which should be run once upon git clone, will install a git post-merge hook into your local check out.

The installed hook will be executed by git automatically at the end of git pull only if it triggered an actual merge event and that the latter was successful.

To trust run:


You don't need to run it if you run:


To distrust run:

rm .git/hooks/post-merge

Validate any notebooks you're contributing to

If you were using a text editor to make changes, when you are done working on a notebook improvement, please, make sure to validate that notebook's format, by simply loading it in the jupyter notebook.

Alternatively, you could use a CLI JSON validation tool, e.g. jsonlint:

jsonlint-php example.ipynb

but it's second best, since you may have a valid JSON, but invalid notebook format, as the latter has extra requirements on which fields are valid and which are not.

Update the documentation notebooks

To update the documentation notebooks to reflect changes in the library, you should use update_notebooks or the tools/update-nbs script. Only use this when you have added a new function that you want to document.

Updating docs from within notebook

update_notebooks can be run from the notebook.

Default - updates all notebooks in the fastai/docs_src folder:


To update specific python file only:

update_notebooks('gen_doc.gen_notebooks.ipynb', update_nb=True)

update_nb=True inserts newly added module methods into the docs that haven't already been documented

Alternatively, you can update a specific module:

update_notebooks('fastai.gen_doc.gen_notebooks', dest_path='fastai/docs_src')

Updating notebooks from script

update_notebooks can be run as a script for convenience.

To update all notebooks under docs_src:

python tools/update-nbs

To automatically append new fastai methods to their corresponding documentation notebook:

python tools/update-nbs --update_nb_links

Update the documentation website

Once notebooks have been updated, update the website html files (under docs)

Updating notebook metadata

In order to pass the right settings to the website version of the docs, each notebook has a custom entry which if you look at the source code, looks like:

 "metadata": {
  "jekyll": {
   "keywords": "fastai",
   "toc": "false",
   "title": "Welcome to fastai"

Do not edit this entry manually, or your changes will be overwritten in the next metadata update.

The only correct way to change any notebook's metadata is by opening docs_src/jekyll_metadata.ipynb, finding the notebook you want to change the metadata for, changing it, and running the notebook, then saving and committing it and the resulting changes.

Updating sidebar

  1. edit docs_src/sidebar/
  2. python tools/
  3. check docs/_data/sidebars/home_sidebar.yml
  4. git commit docs_src/sidebar/ docs/_data/sidebars/home_sidebar.yml

jekyll sidebar documentation.

Generating html

To convert docs_src/*ipynb to docs/*html:

  • only the modified *ipynb:
python tools/build-docs
  • force to rebuild all *ipynbs:
python tools/build-docs -f

After you commit doc changes please validate that all the links and #anchors are correct.

If it's the first time you are about to run the link checker, install the prerequisites first.

After committing the new changes, first, wait a few minutes for github pages to sync, otherwise you'll be testing an outdated live site.

Then, do:

cd tools/checklink

The script will be silent and only report problems as it finds them.

Remember, that it's testing the live website, so if you detect problems and make any changes, remember to first commit the changes and wait a few minutes before re-testing.

You can also test the site locally before committing your changes, please see: README.

To test the site, do:


Working with Markdown


If you work on markdown (.md) files it helps to be able to validate your changes so that the resulting layout is not broken. grip seems to work quite well for this purpose (pip install grip). For example:

grip -b docs/dev/

will open a browser with the rendered markdown as html - it uses github API, so this is exacly how it'll look on github once you commit it. And here is a handy alias:

alias grip='grip -b'

so you don't need to remember the flag.

Markdown Tips

  • If you use numbered items and their number goes beyond 9 you must switch to 4-whitespace chars indentation for the paragraphs belonging to each item. Under 9 or with * you need 3-whitespace chars as a leading indentation.
  • When building tables make sure to use --|-- and not --+-- to separate the headers - github will not render it properly otherwise.

Testing site locally

Install prerequisites:

sudo apt install ruby-bundler

When running this one it will ask for your user's password (basically running a sudo operation):

bundle install jekyll

Start the website:

cd docs
bundle exec jekyll serve

it will tell you which localhost url to go to to see the site.