NumJs is built on top of ndarray and uses many scijs packages. ndarray
is exported as nj.ndarray
, and any NdArray
instance's property selection
is a ndarray
object.
This project is modified from https://github.com/nicolaspanel/numjs and does below modifications
- Remove the feature of images manipulation whose dependencies may result in some installation failure on Mac M1 ARM machine. You could consider ndarray-pixels if you need this feature.
- Add TypeScript typings and
.d.ts
is out of box, JavaScript is supported, too. Also, it includes- ES6 build (ES2015) with CommonJS module for main build in package.json.
- ES6 build (ES2015) with ES6 module for module build. Some tools will follow the module field in package.json, like Rollup, Webpack, or Parcel.
- Refactor internal code via ES6 syntax and does not change the core algorithm code.
- Add "uint8_clamped" (Uint8ClampedArray) support.
- Other improvements.
You can check the changelog.
NumJs is a npm package for scientific computing with JavaScript. It contains among other things:
- a powerful N-dimensional array object,
NdArray
- linear algebra function
- fast Fourier transform
Besides its obvious scientific uses, NumJs can also be used as an efficient multi-dimensional container of generic data.
It works both in node.js and in the browser.
NumJs is licensed under the MIT license, enabling reuse with almost no restrictions.
Try this jsfiddle to play around with the library.
npm install @d4c/numjs
# or
yarn add @d4c/numjs
then using ES6 import:
import nj from "@d4c/numjs";
Or CommonJS require:
// TypeScript users will not get typings when using require
const nj = require('@d4c/numjs').default;
Or download from CDN:
<script type="module">
import nj from 'https://cdn.jsdelivr.net/npm/@d4c/numjs/build/module/numjs.min.js';
// optional: assign to global to let other code can use: globalThis.nj = nj;
</script>
If using ES6 import
or CommonJS require
resuls in some errors, please try to use import nj from "@d4c/numjs/build/module/numjs.min.js
in your environments.
> const a = nj.array([2,3,4]);
> a
array([ 2, 3, 4])
> const b = nj.array([[1,2,3], [4,5,6]]);
> b
array([[ 1, 2, 3],
[ 4, 5, 6]])
Note: Default data container is JavaScript Array
object. If needed, you can also use typed array such as Uint8Array
:
> const a = nj.uint8([1,2,3]);
> a
array([ 1, 2, 3], dtype=uint8)
Below are alternative ways to create same NdArray
.
const a = nj.array([1, 2, 3], "uint8");
const b = nj.array([1, 2, 3], Uint8Array);
const c = nj.array(new Uint8Array([1, 2, 3]));
const d = nj.arange(3,"uint8"); // results in array([ 0, 1, 2], dtype=uint8)
// but we want [1,2,3], how?
// we can manually create a scijs/ndarray object, then assign it as selection property
d.selection = nj.ndarray(new Uint8Array([1, 2, 3]));
// d.selection.data is the stored raw Uint8Array([1, 2, 3])
Note: possible types are int8, uint8, int16, uint16, int32, uint32, float32, float64, uint8_clamped and array (the default)
To create arrays with a given shape, you can use zeros
, ones
or random
functions:
> nj.zeros([2,3]);
array([[ 0, 0, 0],
[ 0, 0, 0]])
> nj.ones([2,3,4], 'int32') // dtype can also be specified
array([[[ 1, 1, 1, 1],
[ 1, 1, 1, 1],
[ 1, 1, 1, 1]],
[[ 1, 1, 1, 1],
[ 1, 1, 1, 1],
[ 1, 1, 1, 1]]], dtype=int32)
> nj.random([4,3])
array([[ 0.9182 , 0.85176, 0.22587],
[ 0.50088, 0.74376, 0.84024],
[ 0.74045, 0.23345, 0.20289],
[ 0.00612, 0.37732, 0.06932]])
To create sequences of numbers, NumJs provides a function called arange
:
> nj.arange(4);
array([ 0, 1, 2, 3])
> nj.arange( 10, 30, 5 )
array([ 10, 15, 20, 25])
> nj.arange(1, 5, 'uint8');
array([ 1, 2, 3, 4], dtype=uint8)
The code is like
// pass data, shape, stride, offset arguments.
// this will call nj.ndarray([2, 3, 4], [3], [1], 0) and assign it to NdArray's selection property
const a = new nj.NdArray([2, 3, 4], [3], [1], 0);
Or you want to apply on a raw scijs ndarray object
// scijs ndarray object
const a = nj.ndarray([2, 3, 4], [3], [1], 0);
NdArray and ndarray
are also exported. You can also use
import { NdArray, ndarray } from "@d4c/numjs"
// or CommonJS:
const NdArray = require('numjs').NdArray;
const ndarray = require('numjs').ndarray;
to reduce typing "nj.
".
NumJs’s array class is called NdArray
. It is also known by the alias array
. The more important properties of an NdArray
object are:
NdArray#ndim
: the number of axes (dimensions) of the array.NdArray#shape
: the dimensions of the array. This is a list of integers indicating the size of the array in each dimension. For a matrix with n rows and m columns, shape will be [n,m]. The length of the shape is therefore the number of dimensions, ndim.NdArray#size
: the total number of elements of the array. This is equal to the product of the elements of shape.NdArray#dtype
: a string describing the type of the elements in the array.int32
,int16
, andfloat64
are some examples. Default dtype isarray
.
An NdArray
can always be converted to a native JavaScript Array
using NdArray#tolist()
method.
Example:
> a = nj.arange(15).reshape(3, 5);
array([[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[ 10, 11, 12, 13, 14]])
> a.shape
[ 3, 5]
> a.ndim
2
> a.dtype
'array'
> a instanceof nj.NdArray
true
> a.tolist() instanceof Array
true
> a.get(1,1)
6
> a.set(0,0,1)
> a
array([[ 1, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[ 10, 11, 12, 13, 14]])
Use nj.array([2,3,4]
as an example. In Node.js, console.log(nj.array([2,3,4])
will print beautified content, array([ 2, 3, 4])
. In browser or using debugger in Node.js, please use console.log(nj.array([2,3,4].toString())
to print its beautified content. toString()
is working in browser/Node.js.
When you print the beautified content of an array, NumJs displays it in a similar way to nested lists, but with the following layout:
- the last axis is printed from left to right,
- the second-to-last is printed from top to bottom,
- the rest are also printed from top to bottom, with each slice separated from the next by an empty line.
One-dimensional arrays are then printed as rows, bidimensionals as matrices and tridimensionals as lists of matrices.
> const a = nj.arange(6); // 1d array
> console.log(a);
array([ 0, 1, 2, 3, 4, 5])
>
> const b = nj.arange(12).reshape(4,3); // 2d array
> console.log(b);
array([[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8],
[ 9, 10, 11]])
>
> const c = nj.arange(24).reshape(2,3,4); // 3d array
> console.log(c);
array([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]],
[[ 12, 13, 14, 15],
[ 16, 17, 18, 19],
[ 20, 21, 22, 23]]])
If an array is too large to be printed, NumJs automatically skips the central part of the array and only prints the corners:
> console.log(nj.arange(10000).reshape(100,100))
array([[ 0, 1, ..., 98, 99],
[ 100, 101, ..., 198, 199],
...
[ 9800, 9801, ..., 9898, 9899],
[ 9900, 9901, ..., 9998, 9999]])
To customize this behaviour, you can change the printing options using nj.config.printThreshold
(default is 7
):
> nj.config.printThreshold = 9;
> console.log(nj.arange(10000).reshape(100,100))
array([[ 0, 1, 2, 3, ..., 96, 97, 98, 99],
[ 100, 101, 102, 103, ..., 196, 197, 198, 199],
[ 200, 201, 202, 203, ..., 296, 297, 298, 299],
[ 300, 301, 302, 303, ..., 396, 397, 398, 399],
...
[ 9600, 9601, 9602, 9603, ..., 9696, 9697, 9698, 9699],
[ 9700, 9701, 9702, 9703, ..., 9796, 9797, 9798, 9799],
[ 9800, 9801, 9802, 9803, ..., 9896, 9897, 9898, 9899],
[ 9900, 9901, 9902, 9903, ..., 9996, 9997, 9998, 9999]])
Single element indexing uses get
and set
methods. It is 0-based, and accepts negative indices for indexing from the end of the array:
> const a = nj.array([0,1,2]);
> a.get(1)
1
>
> a.get(-1)
2
>
> const b = nj.arange(3*3).reshape(3,3);
> b
array([[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8])
>
> b.get(1, 1);
4
>
> b.get(-1, -1);
8
> b.set(0,0,1);
> b
array([[ 1, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8]])
It is possible to slice and stride arrays to extract arrays of the same number of dimensions, but of different sizes than the original. The slicing and striding works exactly the same way it does in NumPy:
> const a = nj.arange(5);
> a
array([ 0, 1, 2, 3, 4])
>
> a.slice(1) // skip the first item, same as a[1:]
array([ 1, 2, 3, 4])
>
> a.slice(-3) // takes the last 3 items, same as a[-3:]
array([ 2, 3, 4])
>
> a.slice([4]) // takes the first 4 items, same as a[:4]
array([ 0, 1, 2, 3])
>
> a.slice([-2]) // skip the last 2 items, same as a[:-2]
array([ 0, 1, 2])
>
> a.slice([1,4]) // same as a[1:4]
array([ 1, 2, 3])
>
> a.slice([1,4,-1]) // same as a[1:4:-1]
array([ 3, 2, 1])
>
> a.slice([null,null,-1]) // same as a[::-1]
array([ 4, 3, 2, 1, 0])
>
> const b = nj.arange(5*5).reshape(5,5);
> b
array([[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[ 10, 11, 12, 13, 14],
[ 15, 16, 17, 18, 19],
[ 20, 21, 22, 23, 24]])
>
> b.slice(1,2) // skip the first row and the 2 first columns, same as b[1:,2:]
array([[ 7, 8, 9],
[ 12, 13, 14],
[ 17, 18, 19],
[ 22, 23, 24]])
>
> b.slice(null, [null, null, -1]) // reverse rows, same as b[:, ::-1]
array([[ 4, 3, 2, 1, 0],
[ 9, 8, 7, 6, 5],
[ 14, 13, 12, 11, 10],
[ 19, 18, 17, 16, 15],
[ 24, 23, 22, 21, 20]])
Note that slices do not copy the internal array data, it produces a new views of the original data.
Arithmetic operators such as *
(multiply
), +
(add
), -
(subtract
), /
(divide
), **
(pow
), =
(assign
) apply elemen-twise. A new array is created and filled with the result:
> zeros = nj.zeros([3,4]);
array([[ 0, 0, 0, 0],
[ 0, 0, 0, 0],
[ 0, 0, 0, 0]])
>
> ones = nj.ones([3,4]);
array([[ 1, 1, 1, 1],
[ 1, 1, 1, 1],
[ 1, 1, 1, 1]])
>
> ones.add(ones)
array([[ 2, 2, 2, 2],
[ 2, 2, 2, 2],
[ 2, 2, 2, 2]])
>
> ones.subtract(ones)
array([[ 0, 0, 0, 0],
[ 0, 0, 0, 0],
[ 0, 0, 0, 0]])
>
> zeros.pow(zeros)
array([[ 1, 1, 1, 1],
[ 1, 1, 1, 1],
[ 1, 1, 1, 1]])
>
To modify an existing array rather than create a new one you can set the copy
parameter to false
:
> ones = nj.ones([3,4]);
array([[ 1, 1, 1, 1],
[ 1, 1, 1, 1],
[ 1, 1, 1, 1]])
>
> ones.add(ones, false)
array([[ 2, 2, 2, 2],
[ 2, 2, 2, 2],
[ 2, 2, 2, 2]])
>
> ones
array([[ 2, 2, 2, 2],
[ 2, 2, 2, 2],
[ 2, 2, 2, 2]])
>
> zeros = nj.zeros([3,4])
> zeros.slice([1,-1],[1,-1]).assign(1, false);
> zeros
array([[ 0, 0, 0, 0],
[ 0, 1, 1, 0],
[ 0, 0, 0, 0]])
Note: available for add
, subtract
, multiply
, divide
, assign
and pow
methods.
The matrix product can be performed using the dot
function:
> a = nj.arange(12).reshape(3,4);
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>
> nj.dot(a.T, a)
array([[ 80, 92, 104, 116],
[ 92, 107, 122, 137],
[ 104, 122, 140, 158],
[ 116, 137, 158, 179]])
>
> nj.dot(a, a.T)
array([[ 14, 38, 62],
[ 38, 126, 214],
[ 62, 214, 366]])
Many unary operations, such as computing the sum of all the elements in the array, are implemented as methods of the NdArray
class:
> a = nj.random([2,3])
array([[0.62755, 0.8278,0.21384],
[ 0.7029,0.27584,0.46472]])
> a.sum()
3.1126488673035055
>
> a.min()
0.2138431086204946
>
> a.max()
0.8278025290928781
>
> a.mean()
0.5187748112172509
>
> a.std()
0.22216977543691244
NumJs provides familiar mathematical functions such as sin
, cos
, and exp
. These functions operate element-wise on an array, producing an NdArray
as output:
> a = nj.array([-1, 0, 1])
array([-1, 0, 1])
>
> nj.negative(a)
array([ 1, 0,-1])
>
> nj.abs(a)
array([ 1, 0, 1])
>
> nj.exp(a)
array([ 0.36788, 1, 2.71828])
>
> nj.tanh(a)
array([-0.76159, 0, 0.76159])
>
> nj.softmax(a)
array([ 0.09003, 0.24473, 0.66524])
>
> nj.sigmoid(a)
array([ 0.26894, 0.5, 0.73106])
>
> nj.exp(a)
array([ 0.36788, 1, 2.71828])
>
> nj.log(nj.exp(a))
array([-1, 0, 1])
>
> nj.sqrt(nj.abs(a))
array([ 1, 0, 1])
>
> nj.sin(nj.arcsin(a))
array([-1, 0, 1])
>
> nj.cos(nj.arccos(a))
array([-1, 0, 1])
>
> nj.tan(nj.arctan(a))
array([-1, 0, 1])
An array has a shape given by the number of elements along each axis:
> a = nj.array([[ 0, 1, 2, 3], [ 4, 5, 6, 7], [ 8, 9, 10, 11]]);
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
> a.shape
[ 3, 4 ]
The shape of an array can be changed with constious commands:
> a.flatten();
array([ 0, 1, 2, ..., 9, 10, 11])
>
> a.T // equivalent to a.transpose(1,0)
array([[ 0, 4, 8],
[ 1, 5, 9],
[ 2, 6, 10],
[ 3, 7, 11]])
>
> a.reshape(4,3)
array([[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8],
[ 9, 10, 11]])
>
Since a
is matrix we may want its diagonal:
> nj.diag(a)
array([ 0, 5, 10])
>
The identity array is a square array with ones on the main diagonal:
> nj.identity(3)
array([[ 1, 0, 0],
[ 0, 1, 0],
[ 0, 0, 1]])
Several arrays can be stacked together using concatenate
function:
> a = nj.arange(12).reshape(3,4)
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>
> b = nj.arange(3)
array([ 0, 1, 2])
>
> nj.concatenate(a,b.reshape(3,1))
array([[ 0, 1, 2, 3, 0],
[ 4, 5, 6, 7, 1],
[ 8, 9, 10, 11, 2]])
Notes:
- the arrays must have the same shape, except in the last dimension
- arrays are concatenated along the last axis
It is still possible to concatenate along other dimensions using transpositions:
> a = nj.arange(12).reshape(3,4)
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>
> b = nj.arange(4)
array([ 0, 1, 2, 3])
>
> nj.concatenate(a.T,b.reshape(4,1)).T
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[ 0, 1, 2, 3]])
> a = nj.array([1, 2, 3])
> b = nj.array([2, 3, 4])
> nj.stack([a, b])
array([[1, 2, 3],
[2, 3, 4]])
> nj.stack([a, b], -1)
array([[1, 2],
[2, 3],
[3, 4]])
Notes:
- the arrays must have the same shape
- take an optional axis argument which can be negative
The clone
method makes a complete copy of the array and its data.
> a = nj.arange(12).reshape(3,4)
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>
> b = a.clone()
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>
> a === b
false
>
> a.set(0,0,1)
> a
array([[ 1, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
> b
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
fft
and ifft
functions can be used to compute the N-dimensional discrete Fourier Transform and its inverse.
Example:
> RI = nj.concatenate(nj.ones([10,1]), nj.zeros([10,1]))
array([[ 1, 0],
[ 1, 0],
[ 1, 0],
...
[ 1, 0],
[ 1, 0],
[ 1, 0]])
>
> fft = nj.fft(RI)
array([[ 10, 0],
[ 0, 0],
[ 0, 0],
...
[ 0, 0],
[ 0, 0],
[ 0, 0]])
>
> nj.ifft(fft)
array([[ 1, 0],
[ 1, 0],
[ 1, 0],
...
[ 1, 0],
[ 1, 0],
[ 1, 0]])
Note: both fft
and ifft
expect last dimension of the array to contain 2 values: the real and the imaginary value
convolve
function compute the discrete, linear convolution of two multi-dimensional arrays.
Note: The convolution product is only given for points where the signals overlap completely. Values outside the signal boundary have no effect. This behaviour is also known as the 'valid' mode.
Example:
> x = nj.array([0,0,1,2,1,0,0])
array([ 0, 0, 1, 2, 1, 0, 0])
>
> nj.convolve(x, [-1,0,1])
array([-1,-2, 0, 2, 1])
>
> const a = nj.arange(25).reshape(5,5)
> a
array([[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[ 10, 11, 12, 13, 14],
[ 15, 16, 17, 18, 19],
[ 20, 21, 22, 23, 24]])
> nj.convolve(a, [[ 1, 2, 1], [ 0, 0, 0], [-1,-2,-1]])
array([[ 40, 40, 40],
[ 40, 40, 40],
[ 40, 40, 40]])
> nj.convolve(nj.convolve(a, [[1, 2, 1]]), [[1],[0],[-1]])
array([[ 40, 40, 40],
[ 40, 40, 40],
[ 40, 40, 40]])
Note: convolve
uses Fast Fourier Transform (FFT) to speed up computation on large arrays.
rot90
> m = nj.array([[1,2],[3,4]], 'int8')
> m
array([[1, 2],
[3, 4]])
> nj.rot90(m)
array([[2, 4],
[1, 3]])
> nj.rot90(m, 2)
array([[4, 3],
[2, 1]])
> m = nj.arange(8).reshape([2,2,2])
> nj.rot90(m, 1, [1,2])
array([[[1, 3],
[0, 2]],
[[5, 7],
[4, 6]]])
mod
(since v0.16.0)
> nj.mod(nj.arange(7), 5)
> m
array([0, 1, 2, 3, 4, 0, 1])