Did you try optimising one of the specific MobileNet models I was using? These should work out of the box. Unfortunately some of the specifics of those models are hardwired into the benchmarking script. You may need to tweak things if you’re using a different model, especially one with different sized input layers.