Incorrect loss normalization in semi-supervised training: Cross-entropy should use labeled_minibatch_size

[shamss778]

I think that the current implementation normalizes both classification and consistency losses by minibatch_size instead of using the appropriate denominators for each loss component. (unless I didnt iiunderstand the paper correctly)

Current code:
class_loss = class_criterion(class_logit, target_var) / minibatch_size

But isnt Cross-entropy losses supposed to be normalized by labeled_minibatch_size (samples that contribute to supervised learning).

class_loss = class_criterion(class_logit, target_var) / labeled_minibatch_size