A JavaScript implementation of the FastText prediction algorithm
The goal is to provide a compatible predict and predict-prob with the C++
version of FastText for use in Node.js.
FastText is a project out of Facebook Research. The primary implementation can be found at https://github.com/facebookresearch/fastText. That is the source code used to create this version.
This code is an experiment to see how fast a pure JavaScript implimentation could be. I envision that a modified version of the original C++ compliled to WebAssembly will be the most performant while having the best interoperability with Node.js.
The original FastText C++ version allows for both the creation of the model via training, as well as the ability to use that model for predictions. This implimentaion is currently only focused on the prediction side of FastText. With that in mind, it is compatible with models generated by the C++ version.
As much as I love Node and JavaScript, I don't see much value in implimenting the training code as the original is multithreaded and the performance will never be matched.
Initial testing with our pretrained models is showing results that are very close to the C++ version. I'm still working through the code to see if I missed something in the port.
I also haven't started preformance tuning the code yet. There are many places for optimizations and doing things the JavaScript way instead of the C++ way.
This is still in progress, but I'd be curious to hear of other's experiences with it.
Several of the values stored in the model file are 64-bit integers, so JavaScript limits apply (max values of 2^53 - 1). Models that are larger than that will probably fail to work correctly.
Currently this code only loads the "compressed" model format (.ftz).
You can find sample models on the FastText website. The one in the test.js file is the
compressed Amazon Review Full which can be found at https://fasttext.cc/docs/en/supervised-models.html#content
fasttextJS is BSD-licensed just like the original source. I can only assume that the patent clause also applies to this port.