A faster Matrix for the Java Script World.

There are pretty awesome matrix libraries for Java Script in the wild. But if you are looking for something faster consider using this library. Matrix-reef.js is a pure Java Script matrix library for numeric computing. It is extremely fast and there is no third-party dependencies under the wood (no webgl!). Check out the comparison below.

This table shows the amount of operations per second achieved by Matrix-reef.js and 3 other popular libraries on multiplying squared matrices:

I got these results on V8/node10.19 @ ubuntu 20.04 with 2-cpu and 8GB RAM over virtualbox.

The performance keeps fast when matrices are not squared:

The label 4x1 means multiply one 4x1 matrix by another 1x4. See Running the benchmarks for more details.

What is the trick? Some libraries are really complete in terms of functionalities, supporting things such as non numerics elements, complex numbers and other very important features. All those requirements seem to come with a price in terms of performance. On the other hand, Matrix-reef.js has only a small - but powerful - subset of features.

const A = new Matrix([
    [1, 2], 
    [3, 4], 
    [5, 6]
]); // create a 3x2 matrix

const B = new Matrix([
    [2, 1], 
    [2, 0]
]); // create a 2x2 matrix

const C = A.multiply(B);

See examples below for a full list of functionalities.

Matrix-reef.js has no third-party, dependency to install.

$ npm install matrix-reef.js

and include

const { Matrix } = require('matrix-reef.js');

<script src="https://cdn.jsdelivr.net/npm/[email protected]/index.min.js"></script>

Matrix-reef.js API is very streamline. Check out the examples below:

• creating: constructors, zeros, ones, identity, diagonal matrix from vector, cloning and slicing.
• accessing: get, set, row, col, diagonal and path.
• arithmetics: add, subtract, add scalar, add row to matrix and add column to matrix.
• multiplication: multiply by scalar, dot multiply and matrix multiplication.
• functional: map, max, min, sum, product, rowWise, colWise and chaining calls.
• matrices: dimension, rows, cols, transpose, determinant, inverse, reshape, flatten, rotate90, rotate180 and rank.
• query: comparing, isSquared, isDiagonal, isUpperTriangular and isLowerTriangular.
• optimization: in-place operations, views, Matrix2D, Matrix3D, Matrix4D, Vector, RowVector, ColumnVector, Vector2D, Vector3D, Vector4D and Sparse Matrix.
• machine learning: convolution, de-convolution, pooling, padding and stribes.
• miscellaneous: print out, convert to array and shuffle.

PS.: Some features are not available yet.

This example shows how to train a neural network using Iris dataset and only add, multiply, dotMultiply, map and transpose. You can run it by executing the following command:

$ node examples/training.ann.js 

THe output should be like:

Elapsed time: 0.81 secs for 5000 epochs.

There are several ways to evaluate a library. Here we are focused on throughput (that is, operations/sec). The code for this benchmark is here. It uses Benchmark.js to compare the performance of Matrix-reef.js and three other Java Script matrix libraries.

To run the benchmark is easy. Just execute the following command to run a benchmark yourself:

node benchmarks/multiplication.js 10 4 10 5

The example above executes chunks of 10 4x10 by 10x5 matrix multiplications.

I'm working hard on this but I'm afraid that Matrix-reef.js is still in the early stages and likely to do not be ready for production yet. My todo list is in the issues section.

Do not hesitate to open a new issue whenever you find something wrong or does not working as expected. Suggestions of new features are pretty welcome as well as collaborations by PR test cases, benchmarks or even new features.