Training dynamic ML models on iOS 15
summary
Shows how iOS 15 lets apps train style‑transfer models on device (example app Demi) by feeding a style image plus ~25 content images into CreateML and exporting CoreML models for fast local inference. Implementation uses CreateML on iOS 15, CoreML, SwiftUI, Combine and async/await and an MLJob with progress publisher; typical runs use ~200–500 iterations, take ~10–15 minutes on a 2019 i7 (inference 1–2s), and produce ~6–7 MB models. Notes: fixed 512×512 input, no simulator support, models grow app storage and are not cross‑platform.
AI-written summary · about
chapters (11)
- 0:00Introduction and speakers
- 4:47User expectations and ML role
- 10:04Server vs device ML trade-offs
- 17:48Core ML and available tools
- 22:45Create ML macOS app overview
- 30:01Style transfer with Create ML
- 39:36Predefined vs dynamic models
- 44:55Training ML models on iOS
- 50:15Building the demo app
- 57:19Code walkthrough and predictions
- 1:00:33Production insights and takeaways
related talks
34mDemystifying ML in iOS
40mNavigating the Machine Learning Landscape for Apple Devices
40mDeep Learning on iOS
22mGetting Started with Apple ML
25mGetting Started with Machine Learning in your iOS apps
44mFederated ML on iOS: Protecting User Data in the Age of AI
10mMachine Learning And React Native
41mTensorFlow Neural Networks on iOS