Hey Eric, great to see you've now published this! I know we chatted about this briefly last year, but it would be awesome to see how the performance of jax-js compares against that of other autodiff tools on a broader and more standard set of benchmarks: https://github.com/gradbench/gradbench
For sure! It looks like this is benchmarking the autodiff cpu time, not the actual kernels though, which (correct me if I’m wrong) isn’t really relevant for an ML library — it’s more for if you have a really complex scientific expression
Nope, both are measured! In fact, the time to do the autodiff transformation isn't even reflected in the charts shown on the README and the website; those charts only show the time to actually run the computations.
Congrats on the launch! This is a very exciting project because the only decent autodiff implementation in typescript was tensorflowjs, which has been completely abandonned by Google. Everyone uses onnx runtime web for inference but actually computing gradients in typescript was surprisingly absent from the ecosystem since tfjs died.
I will be following this project closely! Best of luck Eric! Do you have plans to keep working on it for sometime? Is it a side project or will you abe ble to commit to jax-js longer term?
Yes, we are actively working on it! The goal is to be a full ML research library, not just a model inference runtime. You can join the Discord to follow along
This is really great. I don't do ML stuff. But I some mathy things that would benefit from running in the GPU so it's great to see the Web getting this.
I hope this will help grow the js science community.
I have a project using tfjs and jax-js is very exciting alternative. However during porting I struggle a lot with `.ref` and `.dispose()` API. Coming from tfjs where you garbage collect with `tf.tidy(() => { ... })`, API in jax-js seems very low-level and error-prone. Is that something that can be improved or is it inherent to how jax-js works?
I don’t think tf.tidy() is a sound API under jvp/grad transformations, also it prevents you from using async which makes it incompatible with GPU backends (or blocks the page), a pretty big issue. https://github.com/tensorflow/tfjs/issues/5468
Thanks for the feedback though, just explaining how we arrived at this API. I hope you’d at least try it out — hopefully you will see when developing that the refs are more flexible than alternatives.
Since ONNX is just a model data format, you can actually parse and run ONNX files in jax-js as well. Here’s an example of running DETR ResNet-50 from Xenova’s transformers.js checkpoint in jax-js
I don’t think I intend to support everything in ONNX right now, especially quant/dequant, but eventually it would be interesting to see if we can help accelerate transformers.js with a jax-js backend + goodies like kernel fusion
jax-js is more trying to explore being an ML research library, rather than ONNX which is a runtime for exported models
Readit News
I will be following this project closely! Best of luck Eric! Do you have plans to keep working on it for sometime? Is it a side project or will you abe ble to commit to jax-js longer term?
I hope this will help grow the js science community.
Would `using`[0] help here?
[0]: https://developer.mozilla.org/en-US/docs/Web/JavaScript/Refe...
Thanks for the feedback though, just explaining how we arrived at this API. I hope you’d at least try it out — hopefully you will see when developing that the refs are more flexible than alternatives.
Huggingface’s transformers.js uses it. And I use that for https://workglow.dev (also tensorflow mediapipe though that is using wasm).
I don’t think webnn has gone anywhere and is too restrictive.
https://jax-js.com/detr-resnet-50
I don’t think I intend to support everything in ONNX right now, especially quant/dequant, but eventually it would be interesting to see if we can help accelerate transformers.js with a jax-js backend + goodies like kernel fusion
jax-js is more trying to explore being an ML research library, rather than ONNX which is a runtime for exported models
See: https://caniuse.com/webgpu