JIT/GPU accelerated deep learning for Elixir with Axon v0.1

Yeah, the ability to import external models is great!

Unfortunately, I have to agree. I've been using Elixir for the last year and it's one of the very few languages that I've really enjoyed using, a lot. I would love to see it get traction, but it doesn't have the same momentum as Python, nor Rust for that matter.

I don't think there's anything special about Python and data science/ML, it just happens to be everyone's go-to wrapper language around C code. Personaly, I dislike Python, it has been nothing bit pain for me in the past (not necessarily the language, but the design of popular libraries such as matplotlib), but it's always handy when you need to get something done.

That being said, there are a lot of really high-quality Elixir projects being developed at the moment, Liveview, Livebook (that I think far exceeds Jupyter in innovation), Nx, etc. Perhaps that's one of the benefits of having a smaller community. I like how Elixir has slowly developed into a mature and beautifuly designed language, with focus on stability and not hype.

Yea, only time will tell, but one thing that I love about Elixir is it community, that try to support and improve the "big projects" instead of remaking, eg:

- Phoenix for web dev

- Nerves for IoT / Embedded systems

- Membrane for multimedia / streaming

Hopefully Nx and Axon will be in this list soon :)

> Libraries like this pop up in all kinds of languages to only slowly die off due to lack of community.

I do not agree on the, very specific, subject of deep-learning. Libraries pop up and then die because they lack competitive performances (often, they don't even have GPU capabilities) or are too restrained due to the methods they used to stay performant (restricting the shape of the neural network).

This particular library builds on XLA (the compiler behind JAX) which should give it competitive performance and a flexible enough interface. Once you have that, having a large community is not that important (most deep learning models are built of fairly simple building blocks that are easy to port if need be).

> The ONNX runtime seems to be the only really valuable part.

I actually disagree. :)

If all we had was bindings for ONNX, then we would have to tie ourselves to a single format (and potentially a single runtime). However, having ONNX built on top of a solid foundation (Axon), we get more flexibility and solidity to handle the moving ML landscape.

As an example, I would actually prefer to load pre-trained models from Hugging Face directly into Axon and bypass ONNX altogether. The most immediate benefit for an Elixir developer is that they don't need Python on their machine for the ONNX conversion. But it also means that I can slightly tweak pre-trained models too. And that's another win.

Perhaps your point is that nobody will train a neural network from scratch in Elixir, and maybe that's true, but I would be wary of looking at ONNX as the only possible outcome of the work done so far.

Indeed there is an associated high risk to new efforts like this but that's often part of creating something new. I have heard countless times that there is no value in creating a new programming language, until the value was suddenly there.

There's a library called Nx which abstracts elixir code into backend engines like xla. You could write your own backend too if you wanted.

Edit: xla, not jax

It uses [Nx](https://github.com/elixir-nx/nx/tree/main/nx#readme) (think of it as an equivalent to numpy/jax) which uses Google XLA's as one of the possible backends to compile and run on the GPU.