I mostly used VGG16 but I suppose this more or less translates into inception as well: in VGG16 there are a few CNN layers and two dense layers at the end, the

I mostly used VGG16 but I suppose this more or less translates into inception as well: in VGG16 there are a few CNN layers and two dense layers at the end, the last one having 1000 outputs. The key is in these dense layers. If you want 30x more categories you probably need at least 30x more parameters in these. And I would naively assume that the relationship is more quadratic. The idea is that if the lower layers trained to recognize features that are useful in differentiating 1000 categories, they probably are good enough features to recognize other categories. After all we know that there are eyes detectors there, grid detectors, text detectors, that naturally emerge. ImageNet will continue to increase its number of categories anyway, so unless you have a cluster of TITAN GPU in your basement, you may just as well wait for their results and take their networks.