scikit.js
TypeScript package for predictive data analysis, data preparation and machine learning.
Aims to be a Typescript port of the scikit-learn python library.
This library is for users who wish to train or deploy their models to JS environments (browser, mobile) but with a familiar API.
Generic math operations are powered by Tensorflow.js core layer for faster calculation.
Documentation site: www.scikitjs.org
Installation
Frontend Users
For use with modern bundlers in a frontend application, simply
npm i @tensorflow/tfjs scikitjs
We depend on the tensorflow library in order to make our calculations fast, but we don't ship it in our bundle. We use it as a peer dependency. General usage is as follows.
import * as tf from '@tensorflow/tfjs'
import * as sk from 'scikitjs'
sk.setBackend(tf)
This allows us to build a library that can be used in Deno, Node, and the browser with the same configuration.
Backend Users
For Node.js users who wish to bind to the Tensorflow C++ library, simply import the tensorflow C++ version, and use that as the tf library
npm i @tensorflow/tfjs-node scikitjs
const tf = require('@tensorflow/tfjs-node')
const sk = require('scikitjs')
sk.setBackend(tf)
Note: If you have ESM enabled (by setting type="module" in your package.json), then you can consume this library with import / export, like in the following code block.
import * as tf from '@tensorflow/tfjs-node'
import * as sk from 'scikitjs'
sk.setBackend(tf)
Script src
For those that wish to use script src tags, simply
<script type="module">
import * as tf from 'https://cdn.skypack.dev/@tensorflow/tfjs'
import * as sk from 'https://cdn.skypack.dev/scikitjs'
sk.setBackend(tf)
// or alternatively you can pull the bundle from unpkg
// import * as sk from "https://unpkg.com/scikitjs/dist/web index.min.js"
</script>
Simple Example
import * as tf from '@tensorflow/tfjs'
import { setBackend, LinearRegression } from 'scikitjs'
setBackend(tf)
const lr = new LinearRegression({ fitIntercept: false })
const X = [[1], [2]] // 2D Matrix with a single column vector
const y = [10, 20]
await lr.fit(X, y)
lr.predict([[3], [4]]) // roughly [30, 40]
console.log(lr.coef)
console.log(lr.intercept)
Coming from Python?
This library aims to be a drop-in replacement for scikit-learn but for JS environments. There are some differences in deploy environment and underlying libraries that make for a slightly different experience. Here are the 3 main differences.
1. Class constructors take in objects. Every other function takes in positional arguments.
While I would have liked to make every function identical to the python equivalent, it wasn't possible. In python, one has named arguments, meaning that all of these are valid function calls.
Python
def myAdd(a=0, b=100):
return a+b
print(myAdd()) # 100
print(myAdd(a=10)) # 110
print(myAdd(b=10)) # 10
print(myAdd(b=20, a=20)) # 40 (order doesn't matter)
print(myAdd(50,50)) # 100
Javascript doesn't have named parameters, so one must choose between positional arguments, or passing in a single object with all the parameters.
For many classes in scikit-learn, the constructors take in a ton of arguments with sane defaults, and the user usually only specifies which one they'd like to change. This rules out the positional approach.
After a class is created most function calls really only take in 1 or 2 arguments (think fit, predict, etc). In that case, I'd rather simply pass them positionally. So to recap.
Python
from sklearn.linear_model import LinearRegression
X, y = [[1],[2]], [10, 20]
lr = LinearRegression(fit_intercept = False)
lr.fit(X, y)
Turns into
JavaScript
import * as tf from '@tensorflow/tfjs'
import { setBackend, LinearRegression } from 'scikitjs'
setBackend(tf)
let X = [[1], [2]]
let y = [10, 20]
let lr = new LinearRegression({ fitIntercept: false })
await lr.fit(X, y)
You'll also notice in the code above, these are actual classes in JS, so you'll need to new
them.
2. underscore_case turns into camelCase
Not a huge change, but every function call and variable name that is underscore_case
in python will simply be camelCase
in JS. In cases where there is an underscore but no word after, it is removed.
Python
from sklearn.linear_model import LinearRegression
X, y = [[1],[2]], [10, 20]
lr = LinearRegression(fit_intercept = False)
lr.fit(X, y)
print(lr.coef_)
Turns into
JavaScript
import * as tf from '@tensorflow/tfjs'
import { setBackend, LinearRegression } from 'scikitjs'
setBackend(tf)
let X = [[1], [2]]
let y = [10, 20]
let lr = new LinearRegression({ fitIntercept: false })
await lr.fit(X, y)
console.log(lr.coef)
In the code sample above, we see that fit_intercept
turns into fitIntercept
(and it's an object). And coef_
turns into coef
.
3. Always await calls to .fit or .fitPredict
It's common practice in Javascript to not tie up the main thread. Many libraries, including tensorflow.js only give an async "fit" function.
So if we build on top of them our fit functions will be asynchronous. But what happens if we make our own estimator that has a synchronous fit function? Should we burden the user with finding out if their fit function is async or not, and then "awaiting" the proper one? I think not.
I think we should simply await all calls to fit. If you await a synchronous function, it resolves immediately and you are on your merry way. So I literally await all calls to .fit and you should too.
Python
from sklearn.linear_model import LogisticRegression
X, y = [[1],[-1]], [1, 0]
lr = LogisticRegression(fit_intercept = False)
lr.fit(X, y)
print(lr.coef_)
Turns into
JavaScript
import * as tf from '@tensorflow/tfjs'
import { setBackend, LogisticRegression } from 'scikitjs'
setBackend(tf)
let X = [[1], [-1]]
let y = [1, 0]
let lr = new LogisticRegression({ fitIntercept: false })
await lr.fit(X, y)
console.log(lr.coef)
Contribution Guide
See guide