Cosine-Similarity Loss is missing in the current code

[kikatuso]

In the trainer.py which I assume is used for decoder + [cls] token training, only the flow loss is minimised and the cosine-similarity loss mentioned in the paper is missing. Why?