TorchJS is a JS binding for PyTorch. Its primary objective is to allow running Torch Script inside Node.js program. Complete binding of libtorch is possible but is out-of-scope at the moment.
In test/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, you can load it into Node.js like this
const torch = require("torch-js");
var test_model_path = "test/test_model.pt";
var script_module = new torch.ScriptModule(test_model_path);
console.log(script_module.toString());
var a = torch.rand(1, 5);
console.log(a.toObject());
var b = torch.rand([1, 5]);
console.log(b.toObject());
var c = script_module.forward(a, b);
console.log(c.toObject());
var d = torch.tensor([[0.1, 0.2, 0.3, 0.4, 0.5]]);
console.log(d.toObject());
var e = script_module.forward(c, d);
console.log(e.toObject());
The program above will print something like this on console. Your result will be different from this since a
and b
are random variables.
ScriptModule("/Users/kittipat/torchjs/tests/test_model.pt")
{ data:
Float32Array [
0.5436246991157532,
0.30234378576278687,
0.4031236171722412,
0.8123507499694824,
0.3121740221977234 ],
shape: [ 1, 5 ] }
{ data:
Float32Array [
0.20072013139724731,
0.09114563465118408,
0.588677167892456,
0.14665216207504272,
0.8567551374435425 ],
shape: [ 1, 5 ] }
{ data:
Float32Array [
0.7443448305130005,
0.39348942041397095,
0.9918007850646973,
0.9590029120445251,
1.168929100036621 ],
shape: [ 1, 5 ] }
{ data:
Float32Array [
0.10000000149011612,
0.20000000298023224,
0.30000001192092896,
0.4000000059604645,
0.5 ],
shape: [ 1, 5 ] }
{ data:
Float32Array [
0.8443448543548584,
0.593489408493042,
1.2918007373809814,
1.359002947807312,
1.668929100036621 ],
shape: [ 1, 5 ] }
There are several ways to create tensors
// With TypedArray and shape array
var a = torch.tensor(
new Float32Array([0.1, 0.2, 0.3, 0.4, 0.5]), {
shape: [1, 5],
});
// With array, will create tensor with float32 data type
var b = torch.tensor([
[0.1, 0.2, 0.3],
[0.4, 0.5, 0.6],
]);
// With array and option object
var c = torch.tensor([
[0.1, 0.2, 0.3],
[0.4, 0.5, 0.6],
], {
dtype: torch.float64
});
// With torch.Tensor.fromObject()
var d = torch.Tensor.fromObject({
data: new Float32Array([0.1, 0.2, 0.3, 0.4, 0.5, 0.6]),
shape: [2, 3],
});
This project uses cmake-js
to build Node extension. You will need to download the preview build of libtorch and extract it to an accessible location. The build script assumes you put libtorch
in the same directory as this library. E.g., if you checked out this library to ~/torch-js
, then libtorch
should be at ~/libtorch
. Once you have that, you can run
yarn install
And, to test, run:
node tests/runTorch.js
If it failed to run because libmklml
is missing, you can download it from conda.
conda install libmklml
If conda's lib
directory is in your path, then you should be able to run the command above. Otherwise, you can set environment variable to point to the directory.
On macOS, it would be:
DYLD_LIBRARY_PATH=$CONDA_PREFIX/lib/ node tests/runTorch.js
On Linux, it should be:
LD_LIBRARY_PATH=$CONDA_PREFIX/lib/ node tests/runTorch.js
If there is any error loading the trace file, you might have to resolve it by installing the matching versions of PyTorch and libtorch and regenerate the file.