Hey,
I may be mistaken, but it seems that there is a mistake in the documentation of the loss function NLLLoss in this page:
http://pytorch.org/docs/master/nn.html#nllloss
According to the given explanations for the variables l_n, it seems that in the case where reduce is true, the weight is applied twice for each target loss. Indeed, by using the variables l_ndefined in the case where reduce is false, the weight w_{y_n} is already applied as factor to x_n_{y_n}.