Jason Antic

1/ PSA/rant: If your'e a developer trying to port a deep learning model from Onnx to Core ML and are finding that Apple's onnx-coreml conversion script isn't working out of the box- you aren't alone. While iOS13 has brought about great improvements to what you can do in Core ML,

/7 Also also.... Just put all your image normalization/denormalization code in additional layers sandwiching your model before converting it to Onnx. It's easier to reason about and do correctly where as Apple's code again is strange and not quite right.

Hey there, I’m planning to follow a similar path but wondering if it’s possible to just rebuild the model entirely in swift, so as to fully utilize all of apple’s hardware magic?

You mean like train it and everything else? Honestly I haven’t bothered digging into that because I’m very happy with FastAI and haven’t seen any compelling reason to move off it. I (superficially) think you may be able to but not sure if it would be worth it....

coreml (not swift) is pretty much required if you want to “utilize all of apple’s hardware magic“