Improve MLX Swift by creating frameworks other than MLX Swift

[KC-2001MS]

MLX Swift is an excellent machine learning framework, but one reason it isn't as widely used as Python is the lack of supporting frameworks. Python has well-established auxiliary frameworks like torchvision and torchtext centered around PyTorch, enabling easy handling of popular datasets for training and experimentation.
Similarly, Swift needs frameworks that work in tandem with MLX. For instance, establishing mechanisms to support data acquisition, preprocessing, and model training would significantly enhance the machine learning environment in Swift. Developing such frameworks could bridge the gap with Python and further elevate the value of MLX Swift. Shall we build them?

[awni]

Agreed the ecosystem needs some love. A good way to help is to contribute to existing projects like MLX Swift, the examples repo, Swift Transformers. Or you can build some new missing part of the ecosystem in Swift!

[KC-2001MS]

Thanks for your reply! That makes sense — contributing to existing projects or building new parts of the ecosystem both sound like great ways to help.

I was thinking — it might be helpful to start a discussion (or document) to gather community opinions on which supporting frameworks are most needed for MLX and organize them into a list.

For example:

• MLX Datasets Swift : for easier dataset loading and management
• SwiftMatplot : for visualizing learning curves and images in a GUI

Also, since some tools don’t necessarily require machine learning expertise, perhaps we could involve not only the MLX community but also the broader Swift developer community.
That way, contributors with strong experience in UI frameworks or other kinds of processing could help build these parts of the ecosystem.

What do you think about that approach?