This may be a simple/stupid question. I'm using a recurrent network to predict the next event in a sequence. There are 10 types of events, but they do not occur evenly. It seems like the weights argument to nn.ClassNLLCriterion(weights) is what I should be using, but how do I calculate those weights? I can't decide if it should be the frequency or inverse, e.g. event B occurs 50 times out of 150, is the weight 1/4 or 3/4?