Deep Learning on iOS
summary
Shows how to run CNN inference on-device using MetalPerformanceShaders and Core ML, with models trained in TensorFlow/Keras and exported as binary weights; a demo app recognizes many objects offline (works in airplane mode) after training with ~86 images augmented ~70×, but the custom logo model ends up ~500 MB and uses large memory. Details include MPS layer mappings, Vision for framebuffer handling, Core ML auto-generated classes, and CPU/GPU tradeoffs MPS needs ~2,000 lines for GoogleNet, Core ML simplifies model import, 500 MB custom model, Vision avoids Metal boilerplate.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:22Outline
- 1:50Deep Learning on iOS
- 4:31Demo
- 9:20Implementation
- 15:53Network Implementation
- 17:26Inference
- 19:11Sample App
- 21:54Google Net
- 29:10Deep Running
- 29:43When to choose bananas
- 32:27Question
related talks
25mGetting Started with Machine Learning in your iOS apps
40mNavigating the Machine Learning Landscape for Apple Devices
27mMachine Learning and Mobile
11mAI & Deep Learning Using Core ML and Metal Framework
10mMachine Learning And React Native
01h 03mTraining dynamic ML models on iOS 15
34mDemystifying ML in iOS
38mDeep Learning for detection on a phone