Lux, especially with Enzyme and Reactant is really good for deep learning research. I am aware there is also ongoing work to make SciML solvers compatible with Reactant which would be super interesting as well.
I also read Carroll Vance’s blog post and learned a lot.
I am reading Chollet’s book about Keras3 (the 2025 edition). People would be surprised how far you can go with Keras even for research without having to go down the full PyTorch or JAX way.
As the creator of Pluto.jl said, we should strive to make julia faster and more powerful, but also simpler to learn and use. If we look at how to train a ConvNet or ResNet for image classification on MNIST or CIFAR10 with Lux, the tutorials cover that very well but the code is not exactly straightforward : users have to define very low level functions (some amount of “boilerplate” one could say), write their own training loops, callbacks, etc…
Now if you add GPU, compilation, data movement between processing units, and other things it’s actually hard to get everything right for a beginner friendly training workflow.
Do you think a new package that is a high level training API would be relevant ? What would it look like ?