On-device machine learning: TensorFlow on Android
summary
Demo shows a real Android app doing on‑device inference from the camera, correctly classifying specific candy packages while offline (airplane mode). Workflow: shoot short video per object, extract frames with ffmpeg video frames, retrain last layer of Inception v3 transfer learning on TensorFlow, strip training nodes and quantize with graph transform. Model size drops from ~84 MB to ~21 MB after 8‑bit quantization and pruning, then ships inside or alongside the APK.
AI-written summary · about
chapters (9)
- 0:00Start
- 3:02Train on the server
- 9:29Pictures
- 10:43Sidebar: Convolutional neural networks
- 11:48Training the InceptionV3 model
- 13:23Downsizing with the graph transform tool
- 16:09Load into Android app
- 16:50Supported platforms
- 17:06Community
related talks
31mSupercharging Android Apps with Tensorflow
47mMachine Learning on Android Demystified
15mSimplified Machine Learning on mobile, browsers, and edge devices with on-device ML
46mFire Up Your Smartphone Neurons with On-Device Deep Learning
33mWho’s Afraid of Machine Learning? First Steps with TensorFlow for Android
40mApplied TensorFlow in Android apps
39mMachine Learning on Android Demystified
39mAI/ML on mobile/android with TensorFlow