TensorFlow for Mobile Poets
summary
Setup instructions for Android and iOS, model examples for classification detection and style transfer on-device, and deployment notes for Raspberry Pi are shown with working demos. Claims include an unoptimized model at ~80 MB reduced by quantization to ~7 MB with ~3% accuracy loss, APK overhead of ~12 MB, and advice to use cloud ML engine or serve models from Firebase. Tooling specifics include FFmpeg to extract frames and that TensorFlow Lite is smaller but not yet recommended for production and a Raspberry Pi cucumber sorter that reportedly works like a human.
AI-written summary · about
chapters (20)
- 0:00Start
- 2:47Buzzwords classification
- 12:04Practical Use Cases
- 13:36Cloud vs. Mobile
- 14:31TensorFlow Lite vs. TensorFlow Mobile
- 16:17TensorFlow ecosystem
- 20:24Convolutional Neural Networks (@yufengg)
- 22:51Optimizing model for Mobile (@yufengg)
- 25:33TensorFlow community
- 26:11Building with Android Studio
- 28:41Android Samples
- 30:12TF Detect
- 30:37TF Style
- 32:11Building for iOS
- 32:28iOS Examples: simple
- 32:56iOS Examples: benchmark
- 33:12Raspberry Pi
- 36:04Resources
- 36:55Feedback!
- 37:26Shameless self-promotion
related talks
43mTensorFlow for Mobile Developers
30mTensorFlow for Mobile Heroes
34mTensorFlow Lite
31mTensorflow for Janitors
27mMachine Learning on the edge using TensorFlow Lite
39mFrom TensorFlow to ML Kit: power your mobile application with machine learning
32mAI for Mobile and IoT Devices: TensorFlow Lite
39mLeveraging the power of C++ for efficient machine learning on embedded devices