search

VITA-Group / DeblurGANv2

[ICCV 2019] "DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better" by Orest Kupyn, Tetiana Martyniuk, Junru Wu, Zhangyang Wang
Other
1.01k stars 265 forks source link

Unstable Content_Loss for validation #78

Open NoviceCrom opened 4 years ago

NoviceCrom commented 4 years ago

During the training process, I observed that my Val_G_Loss curve was quite abnormal using RAGAN-LS due to the unstable Val_Content_Loss. All metrics appear normal except that the curve of Validation_Content_Loss metric on tensorboard fluctuates severely. Is there anything I can do to improve its stability? Also, I found the configuration in the paper differs from that in the released code. For example, total trianing epochs, the number of start_epoch when decay starts and the lamda_adv. @Sandbox3aster @t-martyniuk

yfthu commented 4 years ago

I meet the same problem

longint1024 commented 3 years ago

I meet the same problem, too. Have you solved the problem? I found the G-loss is totally random, though the effect of deblurring is getting better.