Validation set can have the same batch size as your training set, Only the final dataset that you need to predict the output for, should have either batch size equal to one or equal to a number that exactly divides your dataset. That’s because if you’re training or checking the accuracy with your validation set,even if your batch size doesn’t matter the generator resamples the already seen data to fit the batch size,this doesn’t really matter here but if you’re predicting the output,then you surely don’t want to be end up with 3 more extra predictions,or 5 more less predictions than the actual data.