This post shows how to use Electron as the GUI component of Python applications. (Updated version of one of my previous posts.) The frontend and backend communicate with each other using zerorpc. The complete code is on GitHub repo.
The following are copied from my original post. They should be the same. If there are inconsistencies, the README.md on the GitHub repo is more accurate.
The current post is a updated version of the previous post a few years before. Readers do NOT need to read the previous post if haven't.
I didn't expect that the previous post attracted so many visitors. Some other people even posted it on Hacker News and Reddit. The previous post also attracted some criticisms. Here I would like to share my replies to some debates.
Do you know Tkinter, GTK, QT (PySide and PyQT), wxPython, Kivy, thrust, ...?
Yes, I know at least their existences and try a few of them. I still think QT is the best among them. BTW, pyotherside is one of the actively maintaining bindings for Python. I am just offering another "web technology oriented" way here.
... And cefpython.
It's more or less be in "lower level" than where Electron is. For example, PySide is based on it.
Correct. Unless some libraries such as numpy are not available in JS. Moreover, the original intention is using Electron / JavaScript / web technologies to enhance Python Applications.
I can use QT WebEngine.
Go ahead and give it a try. But since you are using "web engine", why not also give Electron a try?
Yes. One for JavaScript and one for Python. Unfortunately, Python and JavaScript are dynamic languages, which usually need run-time support.
In the previous post, I showed an example architecture: Python to build up a localhost server, then Electron is just a local web browser.
start
|
V
+------------+
| | start
| +-------------> +-------------------+
| electron | sub process | |
| | | python web server |
| (basically | http | |
| browser) | <-----------> | (business logic) |
| | communication | |
| | | (all html/css/js) |
| | | |
+------------+ +-------------------+
That is just one not-so-efficient solution.
Let's reconsider the core needs here: we have a Python application, and a Node.js application (Electron). How to combine them and communicate with each other?
We actually need an interprocess communication (IPC) mechanism. It is unavoidable unless Python and JavaScript have direct FFI for each other.
HTTP is merely one of the popular ways to implement IPC, and it was merely the first thing came up to my mind when I was writing the previous post.
We have more choices.
We can (and should) use socket. Then, based on that, we want an abstract messaging layer that could be implemented with ZeroMq that is one of the best messaging libraries. Moreover, based on that, we need to define some schema upon raw data that could be implemented with zerorpc.
(Luckily, zerorpc fits our needs here because it supports Python and Node.js. For more general languages support, check out gRPC.)
Thus, in this post, I will show another example using zerorpc for communication as follows, which should be more efficient than what I showed in my previous post.
start
|
V
+--------------------+
| | start
| electron +-------------> +------------------+
| | sub process | |
| (browser) | | python server |
| | | |
| (all html/css/js) | | (business logic) |
| | zerorpc | |
| (node.js runtime, | <-----------> | (zeromq server) |
| zeromq client) | communication | |
| | | |
+--------------------+ +------------------+
Attention: the example could be successfully run on my Windows 10 machine with Python 3.6, Electron 1.7, Node.js v6.
We need the python application, python, pip, node, npm, available in command line. For using zerorpc, we also need the C/C++ compilers (cc and c++ in the command line, and/or MSVC on Windows).
The structure of this project is
.
|-- index.html
|-- main.js
|-- package.json
|-- renderer.js
|
|-- pycalc
| |-- api.py
| |-- calc.py
| `-- requirements.txt
|
|-- LICENSE
`-- README.md
As shown above, the Python application is wrapped in a subfolder. In this example, Python application pycalc/calc.py provides a function: calc(text) that could take a text like 1 + 1 / 2 and return the result like 1.5 (assuming it be like eval()). The pycalc/api.py is what we are going to figure out.
And the index.html, main.js, package.json and renderer.js are modified from electron-quick-start.
First of all, since we already have the Python application running, the Python environment should be fine. I strongly recommend developing Python applications in virtualenv.
Try install zerorpc, and pyinstaller (for packaging). On Linux / Ubuntu we may need to run sudo apt-get install libzmq3-dev before pip install.
pip install zerorpc
pip install pyinstaller
# for windows only
pip install pypiwin32 # for pyinstallerIf properly configured, the above commands should have no problem. Otherwise, please check out the guides online.
Secondly, try to configure the Node.js and Electron environment. I assume that node and npm can be invoked in the command line and are of latest versions.
We need to configure the package.json, especially the main entry:
{
"name": "pretty-calculator",
"main": "main.js",
"scripts": {
"start": "electron ."
},
"dependencies": {
"zerorpc": "git+https://github.com/fyears/zerorpc-node.git"
},
"devDependencies": {
"electron": "^1.7.6",
"electron-packager": "^9.0.1"
}
}Clean the caches:
# On Linux / OS X
# clean caches, very important!!!!!
rm -rf ~/.node-gyp
rm -rf ~/.electron-gyp
rm -rf ./node_modules# On Window PowerShell (not cmd.exe!!!)
# clean caches, very important!!!!!
Remove-Item "$($env:USERPROFILE)\.node-gyp" -Force -Recurse -ErrorAction Ignore
Remove-Item "$($env:USERPROFILE)\.electron-gyp" -Force -Recurse -ErrorAction Ignore
Remove-Item .\node_modules -Force -Recurse -ErrorAction IgnoreThen run npm:
# 1.7.6 is the version of electron
# It's very important to set the electron version correctly!!!
# check out the version value in your package.json
npm install --runtime=electron --target=1.7.6
# verify the electron binary and its version by opening it
./node_modules/.bin/electronThe npm install will install zerorpc-node from my fork to skip building from sources.
(Consider adding ./.npmrc in the project folder if necessary.)
All libraries should be fine now.
If the above installation causes any errors even while setting the electron version correctly, we may have to build the packages from sources.
Ironically, to compile Node.js C/C++ native codes, we need to have python2 configured, no matter what Python version we are using for our Python application. Check out the official guide.
Especially, if working on Windows, open PowerShell as Administrator, and run npm install --global --production windows-build-tools to install a separated Python 2.7 in %USERPROFILE%\.windows-build-tools\python27 and other required VS libraries. We only need to do it at once.
Then, clean ~/.node-gyp and ./node_modules caches as described above at first.
Set the npm for Electron, and install the required libraries.
Set the environment variables for Linux (Ubuntu) / OS X / Windows:
# On Linux / OS X:
# env
export npm_config_target=1.7.6 # electron version
export npm_config_runtime=electron
export npm_config_disturl=https://atom.io/download/electron
export npm_config_build_from_source=true
# may not be necessary
#export npm_config_arch=x64
#export npm_config_target_arch=x64
npm config ls# On Window PowerShell (not cmd.exe!!!)
$env:npm_config_target="1.7.6" # electron version
$env:npm_config_runtime="electron"
$env:npm_config_disturl="https://atom.io/download/electron"
$env:npm_config_build_from_source="true"
# may not be necessary
#$env:npm_config_arch="x64"
#$env:npm_config_target_arch="x64"
npm config lsThen install things:
# in the same shell as above!!!
# because you want to make good use of the above environment variables
# install everything based on the package.json
npm install
# verify the electron binary and its version by opening it
./node_modules/.bin/electron(Consider adding ./.npmrc in the project folder if necessary.)
We want to build up a ZeroMQ server in Python end.
Put calc.py into folder pycalc/. Then create another file pycalc/api.py. Check zerorpc-python for reference.
from __future__ import print_function
from calc import calc as real_calc
import sys
import zerorpc
class CalcApi(object):
def calc(self, text):
"""based on the input text, return the int result"""
try:
return real_calc(text)
except Exception as e:
return 0.0
def echo(self, text):
"""echo any text"""
return text
def parse_port():
return 4242
def main():
addr = 'tcp://127.0.0.1:' + parse_port()
s = zerorpc.Server(CalcApi())
s.bind(addr)
print('start running on {}'.format(addr))
s.run()
if __name__ == '__main__':
main()To test the correctness, run python pycalc/api.py in one terminal. Then open another terminal, run this command and see the result:
zerorpc tcp://localhost:4242 calc "1 + 1"
## connecting to "tcp://localhost:4242"
## 2.0After debugging, remember to terminate the Python function.
Actually, this is yet another server, communicated over zeromq over TCP, rather than traditional web server over HTTP.
Basic idea: In the main process, spawn the Python child process and create the window. In the render process, use Node.js runtime and zerorpc library to communicate with Python child process. All the HTML / JavaScript / CSS are managed by Electron, instead of by Python web server (The example in the previous post used Python web server to dynamically generate HTML codes).
In main.js, these are default codes to start from, with nothing special:
// main.js
const electron = require('electron')
const app = electron.app
const BrowserWindow = electron.BrowserWindow
const path = require('path')
let mainWindow = null
const createWindow = () => {
mainWindow = new BrowserWindow({width: 800, height: 600})
mainWindow.loadURL(require('url').format({
pathname: path.join(__dirname, 'index.html'),
protocol: 'file:',
slashes: true
}))
mainWindow.webContents.openDevTools()
mainWindow.on('closed', () => {
mainWindow = null
})
}
app.on('ready', createWindow)
app.on('window-all-closed', () => {
if (process.platform !== 'darwin') {
app.quit()
}
})
app.on('activate', () => {
if (mainWindow === null) {
createWindow()
}
})We want to add some code to spawn Python child process:
// add these to the end or middle of main.js
let pyProc = null
let pyPort = null
const selectPort = () => {
pyPort = 4242
return pyPort
}
const createPyProc = () => {
let port = '' + selectPort()
let script = path.join(__dirname, 'pycalc', 'api.py')
pyProc = require('child_process').spawn('python', [script, port])
if (pyProc != null) {
console.log('child process success')
}
}
const exitPyProc = () => {
pyProc.kill()
pyProc = null
pyPort = null
}
app.on('ready', createPyProc)
app.on('will-quit', exitPyProc)In index.html, we have an <input> for input, and <div> for output:
<!-- index.html -->
<!DOCTYPE html>
<html>
<head>
<meta charset="UTF-8">
<title>Hello Calculator!</title>
</head>
<body>
<h1>Hello Calculator!</h1>
<p>Input something like <code>1 + 1</code>.</p>
<p>This calculator supports <code>+-*/^()</code>,
whitespaces, and integers and floating numbers.</p>
<input id="formula" value="1 + 2.0 * 3.1 / (4 ^ 5.6)"></input>
<div id="result"></div>
</body>
<script>
require('./renderer.js')
</script>
</html>In renderer.js, we have codes for initialization of zerorpc client, and the code for watching the changes in the input. Once the user types some formula into the text area, the JS send the text to Python backend and retrieve the computed result.
// renderer.js
const zerorpc = require("zerorpc")
let client = new zerorpc.Client()
client.connect("tcp://127.0.0.1:4242")
let formula = document.querySelector('#formula')
let result = document.querySelector('#result')
formula.addEventListener('input', () => {
client.invoke("calc", formula.value, (error, res) => {
if(error) {
console.error(error)
} else {
result.textContent = res
}
})
})
formula.dispatchEvent(new Event('input'))Run this to see the magic:
./node_modules/.bin/electron .Awesome!
If something like dynamic linking errors shows up, try to clean the caches and install the libraries again.
rm -rf node_modules
rm -rf ~/.node-gyp ~/.electron-gyp
npm installSome people are asking for the packaging. This is easy: apply the knowledge of how to package Python applications and Electron applications.
User PyInstaller.
Run the following in the terminal:
pyinstaller pycalc/api.py --distpath pycalcdist
rm -rf build/
rm -rf api.specIf everything goes well, the pycalcdist/api/ folder should show up, as well as the executable inside that folder. This is the complete independent Python executable that could be moved to somewhere else.
Attention: the independent Python executable has to be generated! Because the target machine we want to distribute to may not have correct Python shell and/or required Python libraries. It's almost impossible to just copy the Python source codes.
This is tricky because of the Python executable.
In the above example code, I write
// part of main.js
let script = path.join(__dirname, 'pycalc', 'api.py')
pyProc = require('child_process').spawn('python', [script, port])However, once we package the Python code, we should no longer spawn Python script. Instead, we should execFile the generated excutable.
Electron doesn't provide functions to check whether the app is under distributed or not (at least I don't find it). So I use a workaround here: check whether the Python executable has been generated or not.
In main.js, add the following functions:
// main.js
const PY_DIST_FOLDER = 'pycalcdist'
const PY_FOLDER = 'pycalc'
const PY_MODULE = 'api' // without .py suffix
const guessPackaged = () => {
const fullPath = path.join(__dirname, PY_DIST_FOLDER)
return require('fs').existsSync(fullPath)
}
const getScriptPath = () => {
if (!guessPackaged()) {
return path.join(__dirname, PY_FOLDER, PY_MODULE + '.py')
}
if (process.platform === 'win32') {
return path.join(__dirname, PY_DIST_FOLDER, PY_MODULE, PY_MODULE + '.exe')
}
return path.join(__dirname, PY_DIST_FOLDER, PY_MODULE, PY_MODULE)
}And change the function createPyProc to this:
// main.js
// the improved version
const createPyProc = () => {
let script = getScriptPath()
let port = '' + selectPort()
if (guessPackaged()) {
pyProc = require('child_process').execFile(script, [port])
} else {
pyProc = require('child_process').spawn('python', [script, port])
}
if (pyProc != null) {
//console.log(pyProc)
console.log('child process success on port ' + port)
}
}The key point is, check whether the *dist folder has been generated or not. If generated, it means we are in "production" mode, execFile the executable directly; otherwise, spawn the script using a Python shell.
In the end, run electron-packager to generate the bundled application. We also want to exclude some folders (For example, pycalc/ is no longer needed to be bundled), using regex (instead of glob, surprise!). The name, platform, and arch are inferred from package.json. For more options, check out the docs.
# we need to make sure we have bundled the latest Python code
# before running the below command!
# Or, actually, we could bundle the Python executable later,
# and copy the output into the correct distributable Electron folder...
./node_modules/.bin/electron-packager . --overwrite --ignore="pycalc$" --ignore="\.venv" --ignore="old-post-backup"
## Packaging app for platform win32 x64 using electron v1.7.6
## Wrote new app to ./pretty-calculator-win32-x64I do not check asar format's availability. I guess it will slow down the startup speed.
After that, we have the generated packaged Electron in current directory! For me, the result is ./pretty-calculator-win32-x64/. On my machine, it's around 170 MB (Electron itself occupies more than 84.2 MB). I also tried to compress it, the generated .7z file is around 43.3 MB.
Copy / Move the folder(s) to anywhere or other machines to check the result! :-)
See GitHub electron-python-example.
issue #6: ... failed with KeyError
issue #7: Uncaught Error: Module version mismatch. Expected 50, got 48.
Uninstall everything, set up the npm environment variables correctly especially for the electron version, remember to activate the virtualenv if using Python virtualenv.
Trim some unnecessary files in Python executable by configuring pyinstaller further. Trim Electron (is it possible?). Use even faster IPC methods (though ZeroMQ is one of the fastest in most cases).
What's more, use QT (huh??), rewrite necessary codes in Node.js / Go / C / C++ (huh??). You name it.
Sure. The solution described here can also be applied to any other programming languages besides Python. Except that, if you want to use Electron as GUI of C/C++ applications, I strongly recommend using Node.js native C/C++ communication mechanism instead of using IPC. Moreover, if you have Java, C# application, using Swing or WPF are much more mature choices.
But, unfortunately, Electron is not for mobile applications and it makes little sense even if possible. Please use native GUI on those platforms.
It's still a promising solution. For drawing interface, we want to use some markup language for declarative UI. HTML happens to be one of the best choices, and its companions JS and CSS happen to have one of the most optimized renderers: the web browser. That's why I am (so) interested in using web technologies for GUI when possible. A few years before the web browsers were not powerful enough, but the story is kind of different now.
I hope this post is helpful to you.