Quality difference between CoreML export and General (compact) domain online prediction

Copper Contributor



I'm currently testing using Custom Vision to classify some basic elements like screws vs nuts vs holes. I trained a General (compact) domain model on cropped images as input and it works quite well, when I use the online prediction endpoint. However, when I export the compact model for iOS CoreML, while it still works in general, the quality is much worse for some of the classifiers, meaning it classifies it wrongly. Is this something that is likely in general, i.e. are there quality differences between CoreML and the online version, even when a compact domain is used?

And if so: is there some documentation available on hints on how to get best results from exported models? As it works really well when using the online endpoint, I'm not sure how to optimize the model.


Thanks a lot!

0 Replies