Online Demos With Pyodide

Trying out Python, including NumPy, in the browser

I’ve been doing some research about the “best” way to make online demos. Obviously that’s pretty subjective, but I think a pretty reasonable goal is not having to completely reimplement code I may have written and used in the course of doing my research. That’s a bit tricky because a demo that runs in the browser usually means JavaScript, while research code could be Python / Julia / whatever. In this post I’ll go through one way of easing this disconnect using Pyodide, a project that brings Python to the browser using WebAssembly. I’d be interested to hear of other possibilities. For example there’s a package called NumJs that aims to bring NumPy-powers to JavaScript, but it doesn’t seem that active and maybe that’s just reinventing the wheel?

The easy bit

The slider changes the temperature, with the critical point in the middle. Don't forget to hit pause if your computer starts getting hot.

Here’s a simulation of the 2D Ising model with Glauber dynamics. There are a million of these on the internet of course. What’s a bit unusual about this one is that it is running the following code:


import numpy as np
from pyodide import to_js

class IsingModel:
    def __init__(self, L):
        self.L = L
        self.config = np.random.choice(a=[True, False], size=(L, L))

    def glauber_update(self, beta):
        spins = 2 * self.config - 1
        fields = np.roll(spins, 1, 0) + np.roll(spins, -1, 0) + np.roll(spins, 1, 1) + np.roll(spins, -1, 1)
        delta_E = 2 * spins * fields
        flip_probabilities = 1 / (1 + np.exp(beta * delta_E))
        flips = np.random.random_sample(size=(self.L, self.L)) < flip_probabilities
        # Only update a site with probability 0.5
        flips = np.logical_and(flips, np.random.choice(a=[True, False], size=(self.L, self.L)))
        self.config = np.logical_xor(self.config, flips)
    def to_js(self):
        return to_js(self.config)

In other words, it’s (NumPy assisted) Python. That means I get to use code already written for research, and all that NumPy vectorization goodness.

A clue as to how it works is provided by that line from pyodide import to_js, because you actually can’t get away from JavaScript. This converts a Python object into a JavaScript object, which is then handled by the visualization (more on that in a moment). In our case, it converts a 2D NumPy array of booleans (the states of the Ising spins, with True for spin up and False for spin down) to an Array of JavaScript Arrays. I initially guessed I was going to have to call self.config.tolist() to get a list of lists first, but Pyodide handles the conversion directly. Nice!

The (harder) bit

Hooking that Python code up to the browser involves a bit of glue. The first thing is to load Pyodide itself from a CDN


async function loadSim() {
  let pyodide = await loadPyodide({
  indexURL: "",
  await pyodide.loadPackage("numpy"); 

  // Code has to be loaded into virtual file system
  await pyodide.runPythonAsync(`
  from pyodide.http import pyfetch
  response = await pyfetch("assets/")
  with open("", "wb") as f:
      f.write(await response.bytes())

  return pyodide.pyimport("ising")

loadPyodide returns the pyodide module. If you’re wondering where loadPyodide function came from: it’s loaded as a script in the index.html of the page you’re viewing: we’ll get to that shortly.

Now comes one of the major oddities of using Python in the browser. Usually you import packages and modules from your file system. In the browser you don’t have access to a real file system, so you have to load your code into a “virtual” file system before carrying on (more or less) as normal. In this case, the code at assets/ is written to the virtual file system as, before pyodide.pyimport("ising") does what you’d normally expect import ising to do. So when we call loadSim() it will return our ising module and we’ll be good to go.

For the visualization I’m using p5.js (again loaded in index.html) because it’s very friendly, but you could use whatever visualization library you like, or do it in raw JS if that’s your thing. In p5.js you define functions like setup() and draw() that (you’ve guessd it) setup at the beginning and draw each frame.


let ising_model // will hold the simulation

const w = 5; // Size of each site
let started = false
let paused = false

async function setup() {
    const ising = await loadSim();
    started = true
    const canvas = createCanvas(500, 500);

    const L = floor(width/w);
    ising_model = ising.IsingModel(L) 

    button = createButton('\u23F8');
    button.position(-width, -12, 'relative')
    button.mousePressed(() => isLooping() ? noLoop() : loop());
    slider = createSlider(1, 3.5, 2.269, 0.01);
    slider.parent('ising-simulation')'width', '150px');
    slider.position(0, 20, 'relative').center('horizontal');

Here we call loadSim(), and instantiate our ising_model object, as well as adding a button and slider to control things. Finally, draw() involves calling ising_model.glauber_update(beta) to update the configuration, and then returning it as an array of arrays using ising_model.to_js() before drawing:


function draw() {
  if (started) {  
    ising_model.glauber_update(1 / slider.value())
    config = ising_model.to_js()
    config.forEach((col, colIdx) => {
      col.forEach((site, rowIdx) => {
        if (site) fill(0);
        else fill(255);
        rect(colIdx * w, rowIdx * w, w, w);

Finally, in index.html you need to add the scripts as well as a div that is going to contain the p5.js canvas


<script src = ""></script> 
<script src=""></script>
<script src = "assets/ising.js"></script> 

<div id="ising-simulation" align='center'>

If you want to preview your simulation locally, you’ll have to start a server so the browser can access those local files. With Python that’s as easy as python3 -m http.server.

And that’s it! There’s a bit of boilerplate, but I feel this maybe fits my goal of lowering the barrier between research code and a demo. Pydodide looks like a cool project and I hope it will deliver further performance improvements over time.

Austen Lamacraft
Austen Lamacraft
Professor of Theoretical Physics

Many body physics and machine learning