Jupyter Style Guide#

These guidelines should be followed by notebooks in the documentation. All notebooks in pymc-examples must follow this to the letter, the style is more permissive for the ones on pymc where not everything is available.

The documentation websites are generated by Sphinx, which uses MYST and MYST-NB to parse the notebooks.

Tip

There is a webinar available on contributing to the PyMC example gallery

Template Notebook#

There is a template Jupyter notebook to be used for new notebooks.

General guidelines#

  • Don’t use abbreviations or acronyms whenever you can use complete words. For example, write “random variables” instead of “RVs”.

  • Explain the reasoning behind each step.

  • Attribute quoted text or code, and link to relevant references.

  • Keep notebooks short: 20/30 cells for content aimed at beginners or intermediate users, longer notebooks are fine at the advanced level.

MyST guidelines#

Using MyST allows taking advantage of all sphinx features from markdown cells in the notebooks. All markdown should be valid MyST (note that MyST is a superset of recommonmark). This guide does not teach nor cover MyST extensively, only gives some opinionated guidelines.

  • Never use url links to refer to other notebooks, PyMC documentation or other python libraries documentations. When linking to other notebooks, always use a ref type cross-reference pointing to the target in the First cell.

    Caution

    Using urls links breaks self referencing in versioned docs! And at the same time they are less robust than sphinx cross-references.

    Examples of cross-references

    References to targets within the current project

    That is, notebooks in pymc-examples referring to other notebooks in pymc-examples.

    Pattern:

    {ref}`explicit text <anchor_id>`
    

    Example source:

    {ref}`Kronecker product <GP-Kron>`
    

    Rendered example: Kronecker product

    References to targets of other projects

    Here “other projects” means any sphinx documentation site that was build independently of the current one. Therefore, this includes linking to pymc-examples notebooks from the pymc documentation or vice versa, or linking to other libraries like arviz, numpy, matplotlib…

    Pattern:

    {ref}`explicit text <key:anchor_id>`
    

    Example source:

    {ref}`how to use InferenceData <arviz:working_with_InferenceData>`
    

    Rendered example: how to use InferenceData

    where key in the pattern (arviz in the example) is one of the keys defined in the intersphinx_mapping variable of conf.py such as arviz, numpy, mpl… For the main pymc repo it is located in docs/source/conf.py, for pymc-examples it is in examples/conf.py.

    To identify which anchor_id to use, you need to either look at the source of the document, or use sphobjinv.

    References to python objects

    Pattern

    {type}`import.path`  # to show full import path
    {type}`~import.path`  # to show only object name
    

    where type is func for functions, meth for methods, class for classes, prop for property, etc.

    Example source:

    {class}`~pymc.gp.HSGP`
    

    Rendered example: HSGP

  • If the output (or even code and output) of a cell is not necessary to follow the notebook or it is very long and can break the flow of reading, consider hiding it with a toggle button

  • Consider using Markdown Figures to add captions to images used in the notebook.

  • Use the glossary whenever possible. If you use a term that is defined in the Glossary, link to it the first time that term appears in a significant manner. Use this syntax to add a term reference. Link to glossary source where new terms should be added.

Variable names#

  • Above all, stay consistent with variable names within the notebook. Notebooks using multiple names for the same variable will not be merged.

  • Use meaningful variable names wherever possible. Our users come from different backgrounds and not everyone is familiar with the same naming conventions.

    • Annotate dimensions too. Notebooks are published to be read, so even if the shape is derived from the inputs or you don’t like to use named dims and don’t use them in your personal code, notebooks must use dims, even if annotating and not setting the shape. It makes the code easier to follow, especially for newcomers.

  • Sometimes it makes sense to use Greek letters to refer to variables, for example when writing equations, as this makes them easier to read. In that case, use LaTeX to insert the Greek letter like this $\theta$ instead of using Unicode like θ.

  • If you need to use Greek letter variable names inside the code, please spell them out instead of using unicode. For example, theta, not θ.

  • When using non meaningful names such as single letters, add bullet points with a 1-2 sentence description of each variable below the equation where they are first introduced.

Choosing variable names can sometimes be difficult, tedious or annoying. In case it helps, the dropdown below has some suggestions so you can focus on writing the actual content

Variable name suggestions

Models and sampling results

  • Use idata for sampling results, always containing a variable of type InferenceData.

  • Store inferecedata groups as variables to ease writing and reading of code operating on sampling results. Use underscore separated 3-5 word abbreviations or the group name. Some examples of abbrebiation/group_name: post/posterior, const/constant_data, post_pred/posterior_predictive or obs_data/observed_data

  • For stats and diagnostics, use the ArviZ function name as variable name: ess = az.ess(...), loo = az.loo(...)

  • If there are multiple models in a notebook, assign a prefix to each model, and use it throughout to identify which variables map to each model. Taking the famous eight school as example, with a centered and non_centered model to compare parametrizations, use centered_model (pm.Model object), centered_idata, centered_post, centered_ess… and non_centered_model, non_centered_idata

Dimension and random variable names

  • Use singular dimension names, following ArviZ chain and draw. For example cluster, axis, component, forest, time

  • If you can’t think of a meaningful name for the dimension representing the number of observations such as time, fall back to obs_id.

  • For matrix dimensions, as xarray doesn’t allow repeated dimension names, add a _bis suffix. i.e. param, param_bis.

  • For the dimension resulting from stacking chain and draw use sample, that is .stack(sample=("chain", "draw")).

  • We often need to encode a categorical variable as integers. add _idx to the name of the variable it’s encoding. i.e. from floor and county to floor_idx and county_idx.

  • To avoid clashes and overwriting variables when using pm.Data, use the following pattern:

    x = np.array(...)
    with pm.Model():
        x_ = pm.Data("x", x)
        ...
    

    This avoids overwriting the original x while having idata.constant_data["x"], and within the model x_ is still available to play the role of x. Otherwise, always try to use the same variable name as the string name given to the PyMC random variable.

Plotting

  • Matplotlib figures and axes. Use:

    • fig for matplotlib figures

    • ax for a single matplotib axes object

    • axs for arrays of matplotlib axes objects

    When manually working with multiple matplotlib axes, use local ax variables:

    fig, axs = pyplot.subplots()
    
    ax = axs[0, 1]
    ax.plot(...)
    ax.set(...)
    
    ax = axs[1, 2]
    ax.scatter(...)
    
    fig, axs = pyplot.subplots()
    
    axs[0, 1].plot(...)
    axs[0, 1].set(...)
    
    axs[1. 2].scatter(...)
    

    This makes editing the code if restructuring the subplots easier, only one change per subplot is needed instead of one change per matplotlib function call.

  • It is often useful to make a numpy linspace into an DataArray for xarray to handle aligning and broadcasting automatically and ease computation.

    • If a dimension name is needed, use x_plot

    • If a variable name is needed for the original array and DataArray to coexist, add _da suffix

    Thus, ending up with code like:

    x = xr.DataArray(np.linspace(0, 10, 100), dims=["x_plot"])
    # or
    x = np.linspace(0, 10, 100)
    x_da = xr.DataArray(x)
    

Looping

  • When using enumerate, take the first letter of the variable as the count:

    for p, person in enumerate(persons)
    
  • When looping, if you need to store a variable after subsetting with the loop index, append the index variable used for looping to the original variable name:

    variable = np.array(...)
    x = np.array(...)
    for i in range(N):
        variable_i = variable[i]
        for j in range(K):
            x_j = x[j]
            ...
    

First cell#

The first cell of all example notebooks should have a MyST target, a level 1 markdown title (that is a title with a single #) followed by the post directive. The syntax is as follows:

(notebook_name)=
# Notebook Title

:::{post} Aug 31, 2021
:tags: tag1, tag2, tags can have spaces, tag4
:category: level
:author: Alice Abat, Bob Barceló
:::

The date should correspond to the latest update/execution date, at least roughly (it’s not a problem if the date is a few days off due to the review process before merging the PR). This will allow users to see which notebooks have been updated lately and will help the PyMC team make sure no notebook is left outdated for too long.

Important

The MyST target (the (notebook_name)= bit) is used to link notebooks between each other. It must be notebook specific, for example its file name. Do not copy paste this and leave notebook_name unmodified

Tags can be anything, but we ask you to try to use existing tags to avoid the tag list from getting too long.

Each notebook should have a one or two categories indicating:

  • the level of the notebook (required):

    • beginner (standing crow icon)

    • intermediate (flying dove icon)

    • advanced (dragon icon)

  • the diataxis type (optional for old notebooks):

    • tutorial

    • how-to

    • explanation

    • reference

Authors should list people who authored, adapted or updated the notebook, excluding those who only re-executed a notebook with little to no code or wording changes. Only author names should be added here as this is only metadata of the notebook, self-promotion links and details on the changes should be added at the “Authors” section, see Authorship and attribution for more details.

Extra dependencies#

If the notebook uses libraries that are not PyMC dependencies, these extra dependencies should be indicated together with some advise on how to install them. This ensures readers know what they’ll need to install beforehand and can for example decide between running it locally or on binder.

To make things easier for notebook writers and maintainers, pymc-examples contains a template for this that warns about the extra dependencies and provides specific installation instructions inside a dropdown.

Thus, notebooks with extra dependencies should:

  1. list the extra dependencies as notebook metadata using the myst_substitutions category and then either the extra_dependencies or the pip_dependencies and conda_dependencies. In addition, there is also an extra_install_notes to include custom text inside the dropdown.

    • notebook metadata can be edited from the menu with Edit ‣ Edit notebook metadata

      This will open a window with json formatted text that might look a bit like:

      {
        "kernelspec": {
          "name": "python3",
          "display_name": "Python 3 (ipykernel)",
          "language": "python"
        },
        "language_info": {
          "name": "python",
          "version": "3.9.7",
          "mimetype": "text/x-python",
          "codemirror_mode": {
            "name": "ipython",
            "version": 3
          },
          "pygments_lexer": "ipython3",
          "nbconvert_exporter": "python",
          "file_extension": ".py"
        }
      }
      
      {
        "kernelspec": {
          "name": "python3",
          "display_name": "Python 3 (ipykernel)",
          "language": "python"
        },
        "language_info": {
          "name": "python",
          "version": "3.9.7",
          "mimetype": "text/x-python",
          "codemirror_mode": {
            "name": "ipython",
            "version": 3
          },
          "pygments_lexer": "ipython3",
          "nbconvert_exporter": "python",
          "file_extension": ".py"
        },
        "myst": {
          "substitutions": {
            "extra_dependencies": "bambi seaborn"
          }
        }
      }
      
      {
        "kernelspec": {
          "name": "python3",
          "display_name": "Python 3 (ipykernel)",
          "language": "python"
        },
        "language_info": {
          "name": "python",
          "version": "3.9.7",
          "mimetype": "text/x-python",
          "codemirror_mode": {
            "name": "ipython",
            "version": 3
          },
          "pygments_lexer": "ipython3",
          "nbconvert_exporter": "python",
          "file_extension": ".py"
        },
        "myst": {
          "substitutions": {
            "pip_dependencies": "graphviz",
            "conda_dependencies": "python-graphviz",
          }
        }
      }
      

      The pip and conda specific keys overwrite the extra_installs one, so it doesn’t make sense to use extra_installs if using them. Either both pip and conda substitutions are defined or none of them is.

  2. include the warning and installation advise template with the following markdown right before the extra dependencies are imported:

    :::{include} ../extra_installs.md
    :::
    

Code preamble#

In a cell just below the cell where you imported matplotlib and/or ArviZ (usually the first one), set the ArviZ style to darkgrid (this has to be in another cell than the matplotlib import because of the way matplotlib sets its defaults):

RANDOM_SEED = 8927
rng = np.random.default_rng(RANDOM_SEED)
az.style.use("arviz-darkgrid")

A good practice when generating synthetic data is also to set a random seed as above, to improve reproducibility. Also, please check convergence (e.g. assert all(r_hat < 1.03)) because we sometime re-run notebooks automatically without carefully checking each one.

Reading from file#

Use a try... except clause to load the data and use pm.get_data in the except path. This will ensure that users who have cloned pymc-examples repo will read their local copy of the data while also downloading the data from github for those who don’t have a local copy. Here is one example:

try:
    df_all = pd.read_csv(os.path.join("..", "data", "file.csv"), ...)
except FileNotFoundError:
    df_all = pd.read_csv(pm.get_data("file.csv"), ...)

pre-commit and code formatting#

We run some code-quality checks on our notebooks during Continuous Integration. The easiest way to make sure your notebook(s) pass the CI checks is using pre-commit. You can install it with

pip install -U pre-commit

and then enable it with

pre-commit install

Then, the code-quality checks will run automatically whenever you commit any changes. To run the code-quality checks manually, you can do, e.g.:

pre-commit run --files notebook1.ipynb notebook2.ipynb

replacing notebook1.ipynb and notebook2.ipynb with any notebook you’ve modified.

NB: sometimes, Black will be frustrating (well, who isn’t?). In these cases, you can disable its magic for specific lines of code: just write #fmt: on/off to disable/re-enable it, like this:

# fmt: off
np.array(
    [
        [1, 0, 0, 0],
        [0, -1, 0, 0],
        [0, 0, 1, 0],
        [0, 0, 0, -1],
    ]
)
# fmt: on

Authorship and attribution#

After the notebook content finishes, there should be an ## Authors section with bullet points to provide attribution to the people who contributed to the notebook. The general pattern should be:

## Authors

* <verb> by <author> in <date> ([repo#PR](https://link-to.pr))

where <author> should be the name (multiple people allowed) which can be formatted as a hyperlink to the personal site or GitHub profile of the person, and <date> should preferably be month and year.

The <verb> part should aim to be descriptive of the changes done, for example “updated”, “re-executed”, “authored” or “adapted”, but it is not restricted to anything.

Authors with significant contributions should also be included in the post metadata as indicated in First cell. There are no general and strict guidelines on that, if in doubt, add yourself and ask reviewers for a second opinion. The main reason for that is the authors section here at the bottom aims to be a log of all changes that happen to the notebook, whereas the metadata at the top is used for rendering citation recommendation, so for example, re-executing a notebook that requires no changes is a valuable contribution which will be logged both here in this section and on GitHub but does not match authorship criteria for citing. On the other hand, updating the wording and rendering of a notebook to make it clearer and more friendly to the reader is something that should be added in both places, even if the notebook is not re-executed.

some examples:

## Authors

* Authored by Chris Fonnesbeck in May, 2017 ([pymc#2124](https://github.com/pymc-devs/pymc/pull/2124))
* Updated by Colin Carroll in June, 2018 ([pymc#3049](https://github.com/pymc-devs/pymc/pull/3049))
* Updated by Alex Andorra in January, 2020 ([pymc#3765](https://github.com/pymc-devs/pymc/pull/3765))
* Updated by Oriol Abril in June, 2020 ([pymc#3963](https://github.com/pymc-devs/pymc/pull/3963))
* Updated by Farhan Reynaldo in November 2021 ([pymc-examples#246](https://github.com/pymc-devs/pymc-examples/pull/246))

and

## Authors

* Adapted from chapter 5 of Bayesian Data Analysis 3rd Edition {cite:p}`gelman2013bayesian`
  by Demetri Pananos and Junpeng Lao in June, 2018 ([pymc#3054](https://github.com/pymc-devs/pymc/pull/3054))
* Reexecuted by Ravin Kumar with PyMC 3.6 in March, 2019 ([pymc#3397](https://github.com/pymc-devs/pymc/pull/3397))
* Reexecuted by Alex Andorra and Michael Osthege with PyMC 3.9 in June, 2020 ([pymc#3955](https://github.com/pymc-devs/pymc/pull/3955))
* Updated by Raúl Maldonado in 2021 ([pymc-examples#24](https://github.com/pymc-devs/pymc-examples/pull/24), [pymc-examples#45](https://github.com/pymc-devs/pymc-examples/pull/45) and [pymc-examples#147](https://github.com/pymc-devs/pymc-examples/pull/147))

References#

References should be added to the references.bib file in bibtex format, and cited with sphinxcontrib-bibtex within the notebook text wherever they are relevant.

The references in the .bib file should have as id something along the lines authorlastnameYEARkeyword or libraryYEARkeyword for documentation pages, and they should be alphabetically sorted by this id in order to ease finding references within the file and preventing adding duplicate ones.

References can be cited twice within a single notebook. Two common reference formats are:

{cite:p}`bibtex_id`  # shows the reference author and year between parenthesis
{cite:t}`bibtex_id`  # textual cite, shows author and year without parenthesis

which can be added inline, within the text itself. At the end of the notebook, add the bibliography with the following markdown

## References

:::{bibliography}
:filter: docname in docnames
:::

or alternatively, if you wanted to add extra references that have not been cited within the text, use:

## References

:::{bibliography}
:filter: docname in docnames

extra_bibtex_id_1
extra_bibtex_id_2
:::

Watermark#

watermark is a library which automatically prints the versions of Python and the packages you used to run the NB – reproducibility rocks!

This library should be in your virtual environment if you installed our requirements-dev.txt. Otherwise, run pip install watermark.

First, add a Markdown cell with the ## Watermark title only so it appears in the table of contents. This is the second to last section, above the epilogue/footer. Then, add a code cell to print the versions of Python and packages used in the notebook. This is the last code cell in the notebook.

The p flag is optional (or it may need to have different libraries as input), but should be added if PyTensor or xarray are not imported explicitly. This will also be checked by pre-commit (because we all forget to do things sometimes 😳).

## Watermark
%load_ext watermark
%watermark -n -u -v -iv -w -p pytensor,xarray

Epilogue#

The last cell in the notebooks should be a markdown cell with exactly the following content:

:::{include} ../page_footer.md
:::

The only exception being notebooks that are not on the usual place and therefore need to update the path to page footer for the include to work.


You’re all set now 🎉. You can push your changes, open a pull request, and, once it’s merged, rest with the feeling of a job well done 👏. Thanks a lot for your contribution to open-source, we really appreciate it!