The model was trained with CreateML on the Oxford pet dataset achieving 91% accuracy on test and exported to CoreML format. The model footprint is just 300 KB and inference is only a few ms per image (depending on hardware).
This repo contains example code to show how to perform fast online local inference with a small CoreML model in an iPhone.