how to calculate error in k-fold cross validation

i am implementing 5 fold cross validation using libsvm as classifier. i have a data of 694 members that the class of each member is already defined. but i have to shuffle the data inside each class for lets say 500 times so that each time, a new set of data is created and passed to k-fold. then i have to compute two sets of errors: the overall error and the error per each class. now for the overall error,i am going to have one error point per each round(total number of rounds is f.g 500) and this error point is the sum of the total errors that is achieved at the end of one time of my 5-fold performance. but about the error per class, i don't know exactly what to do. should i calculate the error per each class per each iteration of 5-fold per each round? or like the overall error i should calculate the error per each class as sum of the errors per that class achieved during 5-fold iterations and consider it as the error point of that class per each round? i have 27 classes and i am told to plot 27 different box plots.