Skip to content

Techainer/torch-js

 
 

Repository files navigation

TorchJS

npm version CI publish-prebuild

TorchJS is a JS binding for PyTorch. Its primary objective is to allow running TorchScript inside Node.js program. Complete binding of libtorch is possible but is out-of-scope at the moment.

This fork is maintained by Techainer

Changes after fork

  • Add support for List (Javascript Array), Dict (Javascript Object), String, float (Javascript number) as inputs and outputs.

  • Add CUDA support.

  • Add ops from torchvision.

  • Add async support for forward function.

  • Add async support for call_scripted_function(function_name, **inputs) to call any method of the scripted module.

  • Provide prebuild binary for cross NVIDIA GPU from Pascal to Ampere

  • Fixed some weird conversion arround List[Tensor] input.

  • Updated libtorch to 1.8.1 and torchvision to 0.9.1

Install

To install the forked version, you can install it from npm:

yarn add torch-js@npm:@techainer1t/torch-js

Example

In tests/resources/torch_module.py, you will find the defination of our test module and the code to generate the trace file.

class TestModule(torch.nn.Module):
    def __init__(self):
        super(TestModule, self).__init__()

    def forward(self, input1, input2):
        return input1 + input2

Once you have the trace file, it may be loaded into NodeJS like this

const torch = require("torch-js");
const modelPath = `test_model.pt`;
const model = new torch.ScriptModule(testModelPath);
const inputA = torch.rand([1, 5]);
const inputB = torch.rand([1, 5]);
const res = await model.forward(inputA, inputB);

More examples regarding tensor creation, ScriptModule operations, and loading models can be found in our examples folder.

Languages

  • C++ 44.0%
  • CMake 17.1%
  • JavaScript 13.4%
  • PowerShell 11.2%
  • Shell 8.8%
  • TypeScript 2.8%
  • Other 2.7%