Machine Learning and Mobile
summary
Shows an end to end pipeline that trains a CIFAR convolutional model with Swift for TensorFlow, exports weights via the NumPy Keras bridge to an h5, freezes to a protobuf, converts with MLIR into TensorFlow Lite, and runs the TFLite file on a phone. Demonstrates a live run recognizing a dog at about 70–80% accuracy. Also cites alternatives and benchmarks: Core ML is iOS only, PyTorch mobile exports .pt files, and older MobileNet runs around 30 ms on an older phone.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:21overview
- 1:18assumptions
- 4:45tensorflow-lite
- 5:56pytorch
- 6:51embedded linux
- 8:18custom hardware
- 9:42tensorflow today
- 11:01operator fusion
- 11:41pipelining
- 12:09polyhedral approach
- 12:39glow
- 13:48evolutionary approaches
- 14:51compiler exploration
- 18:57swift/2/mlir/tf-lite
- 24:42recap
- 25:07kudos
- 26:03papers
related talks
29mPushing Machine Learning into your app When TensorFlow and MLKit rule the world
17mDeploy a custom Machine Learning model to mobile
39mMachine Learning on Your Device: The Options
41mTensorFlow Neural Networks on iOS
40mDeep Learning on iOS
39mFrom TensorFlow to ML Kit: power your mobile application with machine learning
30mTensorFlow for Mobile Heroes
27mMachine Learning on the edge using TensorFlow Lite