Skip to content

ansatzcapital/leda

Repository files navigation

leda

Generate static HTML reports with interactive widgets from Jupyter notebooks

PyPI version PyPI Supported Python Versions GitHub Actions (Tests)

Installation

leda is available on PyPI:

pip install leda

Quick Start

Generation

To generate a static HTML report from a Jupyter notebook, run:

python -m leda /path/to/nb.ipynb --output-dir ./outputs/

# Optional args:
python -m leda /path/to/nb.ipynb --output-dir ./outputs/ \
    -i "abc = 123" -k "other_kernel" --cell-timeout 100

This generation automatically includes tweaks to the notebook to make the output look more report-like (e.g., hiding all input code)

See the static demos being served by GitHub Pages.

leda is like:

The -i (--inject) arg is used to inject user code (and set report params) via a new cell prepended to the notebook during generation.

And the --template_name/--theme args allow you to choose between classic, lab (light/dark), and lab_narrow (light/dark).

Note: leda assumes that all code is run in a trusted environment, so please be careful.

Interaction/Widgets

leda also provides an %%interact magic that makes it easy to create outputs based on widgets that work in both dynamic and static modes, e.g.:

# In[ ]:


import leda
import numpy as np
import pandas as pd


# In[ ]:


leda.init("matplotlib")  # Loads `interact` magic when running in Jupyter notebook


# In[ ]:


%%interact column=list("abcdefghij");mult=[1, 2, 3]
df = pd.DataFrame(
  np.random.RandomState(42).rand(100, 10), columns=list("abcdefghij")
)
title = f"column={column!r}, mult={mult}"
(df[[column]] * mult).plot(figsize=(15, 8), lw=2, title=title)

There are two types of interact modes: dynamic and static.

Dynamic mode is when you're running the Jupyter notebook live, in which case you will re-compute the cell output every time you select a different mult. We always use ipywidgets as the dynamic widget backend.

In a static mode (using whichever static widget backend is configured), the library will pre-compute all possible combinations of widget outputs (see Cartesian product) and then render a static HTML report that contains widgets that look and feel like the dynamic widgets (despite being pre-rendered). See below for a list of supported static backends.

Report Web UI Server

Unlike voila, because all report output is static HTML, you can stand up a report web UI server that suits your needs very easily. That means:

  • It's trivial to set up in many cases.
  • It's as scalable as your web server's ability to distribute static content.
  • It's more cost-efficient because there are no runtimes whatsoever.
  • You don't have to worry about old versions no longer working due to code or data changes, so the historical archive of old reports never expires or changes or breaks.

For example, you can generate the report to a file, upload that file to a shared location, and then stand up a bare-bones nginx server to serve the files. Instead of having a two-step process of generation + upload, you could alternatively implement your own leda.ReportPublisher and create a generation script of your own--or use it as a library in client script.

Another example is you can simply host a static S3 bucket, enable website hosting and then either use S3 as a web server publically or via locked down S3 endpoint.

You could also use GitHub Pages, much like the static demos page.

Params

Reports can be parametrized so that the user can set different values for each report run.

In the notebook, just use leda.get_param():

# In[ ]: 


import leda


# In[ ]: 


data_id = leda.get_param("data_id", dynamic_default=1, static_default=2)

And then change the injected code during each run:

python -m leda /path/to/nb.ipynb --output ./outputs/ -i "data_id = 100"

Modular

leda is built to work with multiple visualization and widget libraries.

It works with these visualization libraries:

With the default dynamic widget library:

And with these static widget libraries:

Testing

See the requirements-bundle*.txt for version bundles that we currently test systematically.

Known Issues

  • There are multiple issues using matplotlib with panel, including:
    • The last widget output is not different from the penultimate one: holoviz/panel#1222
    • All the widget outputs show up sequentially, instead of being hidden until chosen. This seems to be a known issue per the panel FAQ; however, using the example fix provided does not work.