deeplearn.js

deeplearn.js is an open-source library that brings performant machine learning building blocks to the web, allowing you to train neural networks in a browser or run pre-trained models in inference mode.

We provide two APIs, an immediate execution model (think NumPy) and a deferred execution model mirroring the TensorFlow API.

deeplearn.js was originally developed by the Google Brain PAIR team to build powerful interactive machine learning tools for the browser. You can use the library for everything from education, to model understanding, to art projects.

Examples

Getting started

deeplearn.js is an open source hardware-accelerated JavaScript library for machine intelligence. deeplearn.js brings performant machine learning building blocks to the web, allowing you to train neural networks in a browser or run pre-trained models in inference mode.

We provide two APIs, an immediate execution model (think NumPy) and a deferred execution model mirroring the TensorFlow API. deeplearn.js was originally developed by the Google Brain PAIR team to build powerful interactive machine learning tools for the browser, but it can be used for everything from education, to model understanding, to art projects.

Usage

yarn add deeplearn or npm install deeplearn

TypeScript / ES6 JavaScript

See the TypeScript starter project and the ES6 starter project to get you quickly started. They contain a short example that sums an array with a scalar (broadcasted):

import {Array1D, NDArrayMathGPU, Scalar} from 'deeplearn';

const math = new NDArrayMathGPU();
const a = Array1D.new([1, 2, 3]);
const b = Scalar.new(2);

const result = math.add(a, b);

// Option 1: With async/await.
// Caveat: in non-Chrome browsers you need to put this in an async function.
console.log(await result.data());  // Float32Array([3, 4, 5])

// Option 2: With a Promise.
result.data().then(data => console.log(data));

// Option 3: Synchronous download of data.
// This is simpler, but blocks the UI until the GPU is done.
console.log(result.dataSync());

ES3/ES5 JavaScript

You can also use deeplearn.js with plain JavaScript. Load the latest version of the library from unpkg:

<script src="https://unpkg.com/deeplearn"></script>

To use a specific version, add @version to the unpkg URL above (e.g. https://unpkg.com/deeplearn@0.2.0), which you can find in the releases page on GitHub. After importing the library, the API will be available as dl in the global namespace.

var math = new dl.NDArrayMathGPU();
var a = dl.Array1D.new([1, 2, 3]);
var b = dl.Scalar.new(2);

var result = math.add(a, b);

// Option 1: With a Promise.
result.data().then(data => console.log(data)); // Float32Array([3, 4, 5])

// Option 2: Synchronous download of data. This is simpler, but blocks the UI.
console.log(result.dataSync());

Development

To build deeplearn.js from source, we need to clone the project and prepare the dev environment:

$ git clone https://github.com/PAIR-code/deeplearnjs.git
$ cd deeplearnjs
$ yarn prep # Installs dependencies.

We recommend using Visual Studio Code for development. Make sure to install TSLint VSCode extension and the clang-format command line tool with the Clang-Format VSCode extension for auto-formatting.

To interactively develop any of the demos (e.g. demos/nn-art/):

$ ./scripts/watch-demo demos/nn-art
>> Starting up http-server, serving ./
>> Available on:
>>   http://127.0.0.1:8080
>> Hit CTRL-C to stop the server
>> 1357589 bytes written to dist/demos/nn-art/bundle.js (0.85 seconds) at 10:34:45 AM

Then visit http://localhost:8080/demos/nn-art/. The watch-demo script monitors for changes of typescript code and does incremental compilation (~200-400ms), so users can have a fast edit-refresh cycle when developing apps.

Before submitting a pull request, make sure the code passes all the tests and is clean of lint errors:

$ yarn test
$ yarn lint

To run a subset of tests and/or on a specific browser (note the extra --):

$ yarn test --browsers=Chrome --grep='multinomial'
 
> ...
> Chrome 62.0.3202 (Mac OS X 10.12.6): Executed 28 of 1891 (skipped 1863) SUCCESS (6.914 secs / 0.634 secs)

To build a standalone ES5 library that can be imported in the browser with a <script> tag:

$ ./scripts/build-standalone.sh # Builds standalone library.
>> Stored standalone library at dist/deeplearn(.min).js

To do a dry run and test building an npm package:

$ ./scripts/build-npm.sh
>> Stored npm package at dist/deeplearn-VERSION.tgz

To install it locally, run npm install ./dist/deeplearn-VERSION.tgz.

On Windows, use bash (available through git) to use the scripts above.

Looking to contribute, and don’t know where to start? Check out our “help wanted” issues.

Supported environments

deeplearn.js targets environments with WebGL 1.0 or WebGL 2.0. For devices without the OES_texture_float extension, we fall back to fixed precision floats backed by a gl.UNSIGNED_BYTE texture. For platforms without WebGL, we provide CPU fallbacks.

While the library supports most devices, our demos don’t currently work on iOS Mobile or Desktop Safari. We are working on updating them, check back soon.

Resources

Thanks

  for providing testing support.

Acknowledgements

deeplearn.js was originally developed by , , and Charles Nicholson.

We would like to acknowledge Chi Zeng, David Farhi, Mahima Pushkarna, Lauren Hannah-Murphy, Minsuk (Brian) Kahng, James Wexler, Martin Wattenberg, Fernanda Viégas, Greg Corrado, Jeff Dean for their tremendous help, and the Google Brain team for providing support for the project.