[Reproducing PhaseNet] Phase Angle not Changing Monotonically

Hi! I am interesting in PhaseNet and I am re-implementing it on my own. I use joint linear velocities in root frame to train the model as stated in the paper, and I got an MSE loss of ~0.007. But the phase angles I get from the model don't look nice: they swing a lot instead of changing monotonically (video attached). What could possibly go wrong here?

https://github.com/user-attachments/assets/a327b2a1-cba2-4cc8-be08-d99013089336