Mobiconf 2018: Meghan Kane, Novoda
summary
Walks through a working pipeline: use Vision rectangle detection to crop a boardgame box, extract video frames via ffmpeg frame extraction, train an image classifier in Create ML on 5 classes (Apple recommends ≥10 images/class) using transfer learning, export an .mlmodel and run it with VNCoreMLModel and VNImageRequestHandler. Reported results show 100% train/validation, 96% test. For style transfer, Turi Create trains a model (example: 300 iterations, ~3 hours on a MacBook Pro) and exports to Core ML.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:57Tour Itinerary
- 1:38ML Tasks
- 4:53Advantages
- 5:27Apple ML Offerings
- 5:52Inference
- 7:30Road trip stops
- 7:41Road Trip Bucket List
- 10:13Why detect the rectangle?
- 10:33Road trip time
- 11:45Vision: detect rectangle
- 14:14Create ML
- 16:13Collecting Data Guidelines
- 18:37Collecting Data Quickly
- 20:10Prepare Data
- 27:05Transfer Learning
- 31:24Vision + Core ML: predict
- 34:02Turi Create
- 34:58Setting up python environment
- 36:01Train style transfer
- 40:30Tour de Apple ML is over..
