Creating smaller, faster, production-worthy mobile machine learning models for Android
summary
Shows that simple prune and quantize steps turn research models into production-ready mobile models: a neural style transfer goes from ~1.6M weights and 7 MB to 11,000 weights and 17 kilobytes from 7 MB, running 30 fps on mid-range phones with indistinguishable quality. Recommends start-with encoders like MobileNet/SqueezeNet/DeepLab, export to TensorFlow Lite, test GPU delegate vs NNAPI per device, and expect up to ~80x runtime variation across Android hardware.
AI-written summary · about
chapters (8)
- 0:00Introduction and background
- 6:56Why mobile ML matters
- 13:11Model selection and training
- 17:21Pruning and quantization techniques
- 21:37Performance optimization tips
- 27:34Real-world examples and demos
- 33:13Getting started with Fritz
- 35:53Q&A on training and hardware
related talks
39mAI/ML on mobile/android with TensorFlow
46mFire Up Your Smartphone Neurons with On-Device Deep Learning
43mTensorFlow for Mobile Developers
14mOptimize your TensorFlow Lite models
39mMachine Learning on Android Demystified
29mRunning Deep Learning Models on Mobile with optimised speed
15mSimplified Machine Learning on mobile, browsers, and edge devices with on-device ML
38mNEURAL NETWORKS ON YOUR MOBILE PHONE. DIVING INTO THE NN API