TensorFlow for Mobile Developers
summary
Shows how to run TensorFlow models on Android, iOS and Raspberry Pi with concrete tradeoffs: Inception v3 (48 layers) yields ~90 MB models that quantize to ~20 MB and can shrink to ~7 MB with 8-bit quantization, APK size increases ~12 MB, and on-device inference enables real-time classification, object counting and style transfer via Magenta. Training often requires GPUs (example benchmark: 1.6 hours on a GPU instance ≈ $10), build needs Bazel 0.6 and Android SDK 23, and models are served either bundled or downloaded on demand. quantize to 8-bit, APK +12 MB, 1.6 h GPU ≈ $10
AI-written summary · about
chapters (7)
- 0:00Introduction to talk
- 11:23Why use ML on mobile
- 20:16Optimizing models for mobile
- 24:12Building with TensorFlow Lite
- 26:16Google’s TensorFlow demos
- 31:34TensorFlow on iOS
- 33:40Resources and closing
related talks
34mTensorFlow Lite
46mTensorFlow for Mobile Poets
30mTensorFlow for Mobile Heroes
19mOn-device machine learning: TensorFlow on Android
40mApplied TensorFlow in Android apps
31mSupercharging Android Apps with Tensorflow
41mTensorFlow Neural Networks on iOS
39mCreating smaller, faster, production-worthy mobile machine learning models for Android