I usually just optimize CNN architectures using Google's intelligence, i.e. let the big companies with infinite compute find good architectures for me.
Also, I wanted to ask: were you using Gooogle Vision, as when I was doing the research it seemed that they do not allow you to export the model.