From TensorFlow to ML Kit: power your mobile application with machine learning
summary
A side project called Magritte recognizes daily objects on-device and speaks vocabulary; it uses transfer learning on TensorFlow, converts to TensorFlow Lite, and moves toward ML Kit with Firebase model hosting. Initial frozen graphs make APKs ~80 MB; post-training quantization cuts size ~4x and MobileNet (used with ~20% of original weights) keeps accuracy while cutting latency to near‑instant on modern phones. TF Lite conversion breaks pre‑1.9, ML Kit is beta with occasional API churn and small runtime overhead.
AI-written summary · about
chapters (28)
- 0:00Intro
- 1:35Amazon Web Services
- 3:16Android Wear 2.0 Smart Reply
- 4:47Machine Learning
- 8:35Here's a Neuron
- 9:08Here's a Neural Network
- 9:48Back Propagation
- 11:01Deep Convolutional Neural Network
- 12:11Transfer Learning
- 13:18Gather Training Data
- 15:05Obtain the Retrained Model
- 15:58Why model size is heavy?
- 16:20Post-training Quantization
- 18:37Optimize for Mobile
- 20:01Model Size Comparison
- 20:15Accuracy Comparison
- 20:25Image Sampling
- 21:05Dispensing Model
- 22:12What's ML Kit for Firebase?
- 26:45Obtain the TF Lite model
- 27:33Input Dimensions & Constants
- 27:46Create local model source
- 27:59Register model sources
- 28:31Input/output options (Float Type)
- 31:39TensorFlow & MobileNet
- 32:42ML Kit: Current State
- 33:46ML Kit: the best is yet to come
- 34:15Federate Learning
related talks
29mFrom TensorFlow to ML Kit: power your Android application with ML
29mPushing Machine Learning into your app When TensorFlow and MLKit rule the world
24mML Kit: Machine learning for mobile developers
38mML Kit: Machine Learning for Mobile with Firebase
21mGetting the best of on-device ML with ML Kit
10mMachine Learning for mobile
30mMachine Learning in Android Apps with ML Kit
04mTop 5 reasons to add ML Kit to your Android or iOS app