Convergence trend of spectrum loss

[RongerSnow]

Thank you for exploring this interesting constraint. In a similar way to yours, I also add spectrum loss to my generated network, but it seems that its value fluctuates all the time. Unlike MSE, it doesn't have a significant downward convergence trend. I am quiet confused whether this is a normal phenomenon? It would be better if you could provide your training log for reference.

[RicardDurall]

Based on our experiments, we have also observed that the spectrum loss in some scenarios can fluctuate. However, as with many constraints, it is a matter of fine-tuning the hyper-parameters.
On our supplemental material, we provide a training log plot of the AE case. There, you can see how exists a certain downward convergence trend (especially in the beginning) and then it flattens. I hope I have helped you.