nicolaspanel/numjs

Like NumPy, in JavaScript

Summary

numjs is a JavaScript port of NumPy's core ideas, providing an N-dimensional array object (NdArray) along with linear algebra, universal math functions, FFT, convolution, and basic image processing. It works in both Node.js and the browser, is MIT-licensed, and builds on the scijs ndarray package.

What it is useful for

The library is useful when you need to do numerical work without leaving JavaScript. Concrete scenarios the README supports:

• Multi-dimensional data handling in Node.js, with typed dtypes (int8, uint8, int16, uint16, int32, uint32, float32, float64) or the default JS Array.
• Browser image pipelines via nj.images.read, nj.images.write, and nj.images.resize, with images represented as [H,W] or [H,W,3] arrays.
• Signal processing using nj.fft, nj.ifft, and nj.convolve (the README notes convolve uses FFT internally for large arrays).
• Element-wise math functions like nj.exp, nj.softmax, nj.sigmoid, nj.sin, and nj.sqrt, plus reductions such as .sum(), .mean(), .std(), .min(), .max().

It is not a drop-in for high-performance scientific Python: the README makes no performance claims and the default storage is plain JS arrays.

How engineers can use it

Install on Node.js (documented):

npm install numjs

var nj = require('numjs');

Install in the browser (documented):

bower install numjs

<script src="https://cdn.jsdelivr.net/gh/nicolaspanel/[email protected]/dist/numjs.min.js"></script>

Working with arrays:

// Create a 3x4 matrix from a range and reshape
var a = nj.arange(12).reshape(3, 4);

// Matrix multiplication
nj.dot(a.T, a);

// Element-wise arithmetic returns a new array
var ones = nj.ones([3, 4]);
ones.add(ones); // [[2, 2, 2, 2], ...]

// Slice with [start, stop, step] semantics
var b = nj.arange(25).reshape(5, 5);
b.slice(null, [null, null, -1]); // reverse each row, like b[:, ::-1]

Image processing:

var img = nj.images.data.lenna; // built-in sample
var resized = nj.images.resize(img, 14, 12);

Documented limitations to be aware of:

• Slices are views, not copies. Use .clone() for an independent array.
• convolve only supports "valid" mode, where signals fully overlap; values outside the signal boundary are ignored.
• fft and ifft expect the last axis to be size 2 (real and imaginary components).
• concatenate requires arrays with matching shape outside the joining axis.

The README points to the numjs globals and NdArray documentation for the full API.