As title suggests I have a modified mnist dataset with multiple digits per image 0-3 and each digit can be 0-9.
I have modelled it as a multi label classification problem and am using cross entropy on 9*3 labels for each sample.
Is this the right way to go about?
My training accuracy is going down but validation accuracy is monotonically increasing
I think I am supposed to do something like cross entropy among the 10 labels for each digit seperately and average them out or something
But I am not able to formulate it.