It mentions allowing you to use additionally trained model is an enhancement that is being worked on for the future!
Still, I think it’s fantastic that you can take a great model in python and make it available in the Elixir machine learning framework, allowing Elixir programs to integrate with this type of models at a low level without completely reinventing all the effort to train and build them in the first place!
Totally agree about ONNX, without it I wouldn't probably even started making this library. However, something I want to try is to code the network with Nx/Axon, without using onnx. The main advantage I see moving from onnx to Axon is to take advantage of all the new cool stuff like Nx sharding: https://dockyard.com/blog/2024/11/06/2024/nx-sharding-update...