Open zxx01 opened 10 months ago
In our experiments, we also sometimes encounter such instability issues, and this unstable training process in INN-based models has been mentioned in other literature as well. In this paper, our solution is to adjust the weights of the message loss and PSNR loss once the accuracy exceeds 95% or nan occurs. Specifically, we reload the model parameters from before the crash and, in the new training session, decrease the weight of the message loss and increase the weight of the PSNR loss.
OK, Thanks!
这是怎么回事?