Well for a while I was lulled into complacency because the retrained networks would indicate 98%+ accuracy, but really that was just an artifact of my 49:1 nothotdog:hotdog image imbalance. When I started weighing proportionately, a lot of networks were measurably lower, although it’s obviously possible to get Inception of Vgg back to a “true” 98% accuracy given enough training time.

That would have beat what I ended up shipping, but the problem of course was the size of those networks. So really, if we’re comparing apples to apples, I’ll say none of the “small”, mobile-friendly neural nets (e.g. SqueezeNet, MobileNet) I tried to retrain did anywhere near as well as my DeepDog network trained from scratch. The training runs were really erratic and never really reached any sort of upper bound asymptotically as they should. I think this has to do with the fact that these very small networks contain data about a lot of ImageNet classes, and it’s very hard to tune what they should retain vs. what they should forget, so picking your learning rate (and possibly adjusting it on the fly) ends up being very critical. It’s like doing neurosurgery on a mouse vs. a human I guess — the brain is much smaller, but the blade says the same size :-/