1.0.0-alpha.4 • Published 5 years ago

libtorchjs v1.0.0-alpha.4

Weekly downloads
9
License
BSD-3-Clause
Repository
github
Last release
5 years ago

LibtorchJS

Simple Node.js N-API module that wraps few pieces of pytorch C++ library (libtorch) to allow loading and running pytorch-trained models in Node.js.

Node.js non-blocking model is great for scheduling heavy computational tasks such as NN inference. This wrapper provides async methods that do not block event loop, so it's possible to use it in a web-service.

This module was made just for fun, to check if it's easily possible to create image style transfer app with Node.js.

Installation

With npm: npm i --save libtorchjs

Module binary is pre-built and published using node-pre-gyp so it's not required to download libtorch/pytorch or install build tools.

Currently, Linux & Windows builds are available.

Usage Example

const torch = require('libtorchjs');

const input = torch.randn([1, 3, 224, 224]);
torch.load('model.pt', function(err, model) {
    model.forward(input, function(err, result) {
        const output = result.toUint8Array();
        console.log(output);
    });
});

API

The overall goal is to mirror pytorch API where possible. Currently just a few methods are exposed.

Tensor

ones(Array shape)

Create tensor of specified shape filled with 1's (autograd is disabled) and return Tensor object.

randn(Array shape)

Create tensor of specified shape (autograd is disabled) filled with random values in (0..1) range.

ScriptModule

load(String filename, callback)

Load traced model async from file and return resulting ScripModule.

forward(Tensor tensor, callback)

Forward tensor async and return resulting Tensor.

Acknowledgments

Following resources were extremely useful for creating this module:

Possible Future

  • Expose more of libtorch
  • Promisify libtorchjs API
  • GPU support
  • Windows :heavy_check_mark:, Mac support
  • Use travis/appveyor for automatic builds