I would like to use the Nengo Loihi CIFAR-10 example from the website and modify it to accommodate CIFAR-100 with its super classes and regular classes. My goal is to compare performance of the Loihi simulation between CIFAR-10 and CIFAR-100. But I am still relatively new to coding in Nengo and Nengo Loihi.
I realize the load command will be changed, as well the need to add the superclass along with updating the regular classes. But other than that, what else would need to be added, updated, or changed?
I have not worked with the CIFAR-100 dataset myself, but looking at the description of it, the image sizes between the two datasets is the same, which should theoretically mean that a network designed for CIFAR-10 will work for CIFAR-100 with minimal changes (if at all).
One change that might need to be performed is to insert additional layers into the network to do the superclass classification, but having not worked with the CIFAR-100 dataset myself, I cannot comment on what structure these additional layers would take. Perhaps @Eric or @drasmuss may have ideas.
It depends a bit what you want to do with it. When I’ve done CIFAR-100 in the past, I’ve just treated it as a straightforward multiclass classification problem with 100 classes. I think this is what people typically do when they report CIFAR-100 results. So for that, the main thing that needs to change with the network is just the output dimension (100 classes instead of 10). That said, you may want a larger/different network architecture to improve performance.
There would be other ways of setting up the classification problem to actually take advantage of the fact that you have both superclass and regular labels, as you mention. I haven’t played around with that at all, though, so I don’t really have any suggestions there.