tch-js

An unofficial JavaScript and TypeScript port of PyTorch C++ library (libtorch).

Install

Package publishing to NPM is still work-in-progress. In the meantime, you can install the package directly from GitHub.

$ npm i git+https://github.com/cedrickchee/tch-js.git

The package will download the pre-built binary during installation. You don't have to download PyTorch/libtorch or install tools for compiling from source.

Versions Supported

• Node.js:
  • 10 (tested)
  • 12 (tested)
  • 14 (WIP)
• PyTorch (CPU):
  • 1.4.X (tested)
  • 1.5.X
  • 1.6.X
  • 1.7.X (tested)

Code Examples

// This is a real example from an audio source separation model.
const { tch, load, Tensor } = require('tch-js');
const fs = require('fs');
const wav = require('node-wav');

// WAV samples of length 269973
const monoAudioChan = new Float32Array([
  1.100000023841858, 1.590000033378601,
  2.049999952316284, 0.18000000715255737
]);
// Flat tensor
let audio = tch.tensor(monoAudioChan); // tensor of size [269973]

// Reshape to 1xSample Length to match model input
audio = audio.view([1, monoAudioChan.length]); // tensor of size [1, 269973]

// Load PyTorch traced model async from file and return resulting ScripModule.
const model = await load('sound-model.pt');
// Forward tensor async and return resulting Tensor.
model.forward(audio, getResult);
const getResult = (err: Error, result: Tensor) => {
  if (err) return;

  // result is a tensor of size [1, 1, 269973]
  const out = result.toFloat32Array(); // convert Tensor to JS TypedArray

  // Encode output to 16-bit float WAV and write to file.
  const buf = wav.encode([out], { sampleRate: 44100, float: false, bitDepth: 16});
  fs.writeFileSync("out.wav", Buffer.from(buf));
};

Build It Yourself

Currently, only Linux builds are available.

If you want to build the package youself, below are the steps to reproduce the build.

Installing on Linux:

• Ubuntu 18.04 (tested)
• Ubuntu 20.04 (tested)

1. Install build tools

$ apt install -y cmake make gcc-c++ unzip

2. Install Node.js 10
3. Download libtorch Download libtorch pre-cxx11 ABI and CPU version.

$ curl -o libtorch.zip https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-1.7.1%2Bcpu.zip
$ unzip libtorch.zip

4. Install Node.js packages

$ npm i --ignore-scripts

5. Build

$ npm run pre-build

6. Test

$ npm run test

The Plan

The library should:

• expose more libtorch types and APIs for inference
• supports Windows
• auto build binaries using CI
• TypeScript types

Research

• PyTorch 1.0 tracing JIT and LibTorch C++ API to integrate PyTorch into NodeJS
• tch-rs - Rust bindings for the C++ API of PyTorch
• torch-js - Node.js binding for PyTorch
• pytorchjs - Torch and TorchVision, but for NodeJS
• Using the PyTorch C++ Frontend tutorial
• Loading a TorchScript Model in C++
• Torchaudio.load() in C++
• Minimal PyTorch 1.0 -> C++ full example demo
• Convert libtorch tensor into an array
• JavaScript TypedArray
• Napi::TypedArray - Node.js N-API doc
• How to use WebAssembly from node.js? - C++ to WebAssembly to Node.js
• node-pre-gyp - Node.js tool for easy binary deployment of C++ addons