We are using TensorFlow.js right as our machine-learning backend to enable browser usage. However, it has many limitations and is almost unmaintained.
After studying alternatives, it seems Burn is the most promising alternative, in terms of current popularity, development pace, existing features, performance and ability to cross-compile to many different targets.
Transitioning means rewriting discojs (about 60% of the repo's typescript code, or slighly under 15k lines) in Rust using this new framework, then re-plugging it into the other repos that use it.
Since there are no pressing deadlines, this is a good time to work on this, and it will make solving other high-priority issues faster once it's done.
We are using TensorFlow.js right as our machine-learning backend to enable browser usage. However, it has many limitations and is almost unmaintained.
After studying alternatives, it seems Burn is the most promising alternative, in terms of current popularity, development pace, existing features, performance and ability to cross-compile to many different targets.
Transitioning means rewriting
discojs(about 60% of the repo's typescript code, or slighly under 15k lines) in Rust using this new framework, then re-plugging it into the other repos that use it.Since there are no pressing deadlines, this is a good time to work on this, and it will make solving other high-priority issues faster once it's done.