I have 22 classes of objects but they have very skewed distributions where max class has 100.000 images and the min class has 1600 images. In that setting I would like to hear some possible solutions to this balance problem.
I have tried followings so far;
multiply number of instances in the lower classes up to the max class by replicating the instances, possibly adding some noise as well.
changing the learning regarding the class distribution in the given minibatch of the next epoch. (no implemented but in my mind)
What are your suggestions?