Semi-Supervised Learning

Modify our loss functions to avoid missing samples

We need at least one ground truth label for each label.