How to Run PyMC in the Browser with PyScript and Pyodide

Congratulations! You're about to be one of the first people who have sampled a [PyMC](https://pymc.io) model in their browser. Either open the webapp or use it here directly:

How is that possible? PyScript!

PyScript is a new (as of 2022) tool that allows you to execute Python directly in your browser. Previously, running code client-side was only possible with JavaScript.

What's really exciting is that it's not just a small subset of Python, but everything. You can even import packages like NumPy, Pandas, and Matplotlib. The way this works is via Pyodide, a port of the CPython runtime implemented in WebAssembly.

If you want to learn more, watch Peter Wang's PyCon 2022 Keynote with many demos.

Possible to run PyMC?

Naturally, I was curious if it was possible to run PyMC through PyScript. On first thought this might seem impossible because PyMC compiles the model evaluation code to C or JAX (through PyTensor). However, PyTensor also has a Python mode which, while being much slower, is fully functional.

Why?

Before, you could have a PyMC model run on the server and then send the results back to the client (i.e. the browser). However, this has a few short-comings:

It's challenging to set everything up to handle the interplay between client and server correctly, with many different technologies interacting in complex ways
On top of the server<->client interplay, you have to pay special attention to scaling as many users might be running compute-extensive PyMC models in parallel
Users might not be comfortable to send their data to your server

If we can just run PyMC in the browser directly, all these problems go away. There is no interplay between client and server because everything runs on the client. There are no scaling issues because users use their own CPUs to fit their models. And finally, no data ever gets transmitted to the server, so it's completely safe and privacy preserving.

The Process

1. Getting PyMC installed in the browser

In PyScript it's possible to install any packages that are on PyPI using micropip.

    <py-config>
      packages = ["pytensor", "pymc", "bokeh"]
    </py-config>

This installs bokeh and pymc in your browser and we can import pymc as pm. Easy-peasy.

2. Write model

Next, we can embed our Python code in py-script tags:

<py-script>
import json
from js import Bokeh, JSON
from bokeh.embed import json_item
from bokeh.plotting import figure

import arviz as az
# Make arviz use bokeh for interactive plotting
az.rcParams["plot.backend"] = "bokeh"

import pymc as pm

def run_model(n=10, k=5):
    # Define model
    with pm.Model() as model:
        p = pm.Beta("p", alpha=1, beta=1)
        obs = pm.Binomial("obs", p=p, n=n, observed=k)
        idata = pm.sample()

    # Generate plot
    p = figure(width=500, height=400, toolbar_location="below")
    az.plot_posterior(idata, var_names=["p"], show=False, ax=p)
    p_json = json.dumps(json_item(p, "myplot"))
    Bokeh.embed.embed_item(JSON.parse(p_json))
</py-script>

Note that because arviz (what PyMC uses for plotting), has support for bokeh, a Python-to-JS plotting library, we can also get interactive plots.

3. That's it!

There is no 3, you just open the website in your browser, it will install the packages and that's it!

I was surprised by how simple it was to get this going, it took me a couple of hours to put everything together. These are really interesting times we're living in.

Applications

So what could we do with this? Well, the possibilities are endless. The main applications will resolve around two possibilities:

Sampling a PyMC model based on user-data
Having a presampled model that we use to make predictions based on user-data

Some example applications could be:

JupyterLite is a Jupyter NB that runs completely in the browser. So you can interactively work with PyMC without having to install anything. This would make our corporate workshops even simpler. Update: This already works.
One of our clients at PyMC Labs perform adaptive psychometric testing. The most informative questions are chosen using a Bayesian model. Currently, there are two versions of the model, one in PyMC that is fit infrequently on a batch of data and one in JavaScript for running in the browser while the subject is doing the test. In the future, they can just use the same PyMC model and don't have to have two separate versions.
Bambi allows to build generalized linear models with a single line that specifies the model. This could easily be turned into a webapp that allows users to upload their data and fit hierachical linear models to it.
Alex Andorra runs a website for electoral forecasting using PyMC. See here for details on the model. Currently it just displays the latest plots generated on the backend, but with this it would allow for custom, user-defined forecasts.
With HelloFresh we have built a state-of-the-art Bayesian marketing mix model. They currently have to run this in Jupyter NB locally and send the results to stake-holders by email. This would make developing a webapp to interact with the data much simpler.

Summary

This feels like a new dawn. JavaScript is by far the most commonly used programming language on the planet. Not because it's a great language (it's OK) but because you can execute it on everything that can run a browser.

This universality is now coming to Python, giving web programmers access to its rich ecosystem, including the PyData stack. And with this blog post, you can also run complex Bayesian models in PyMC.

I cannot wait to see what amazing things the community will produce around this!

Running PyMC in the Browser with PyScript