sanghyun-son / EDSR-PyTorch

PyTorch version of the paper 'Enhanced Deep Residual Networks for Single Image Super-Resolution' (CVPRW 2017)
MIT License
2.44k stars 671 forks source link

GAN Training Issue #258

Open avinabsaha opened 4 years ago

avinabsaha commented 4 years ago

Hi,

I was trying to train a super-resolution model with L1+ Adversarial Loss. The weights I set for L1 is 0.85 and Adversarial Loss is set to 0.15. Any formal way to set it?

Also, I observe as the training progresses, The L1 loss saturates, The GAN loss increases, The Discriminator Loss decreases to a very small value. And the validation PSNR fluctuates. Am I doing something wrong? I think all losses should go down right?

My Super Resolution network is very small ~2000 MACs. Do I need to change the discriminator for a small SR network?

HaolyShiit commented 4 years ago

I have the same question. The L1 loss and GAN loss of generator increase dramaticlly , and the GAN loss of discriminator drops to a small value. The results are very bad. I can't figure out why the issue happens.
If you find the solution, please let me know, thanks!

machlea commented 3 years ago

My GAN loss increases while dis loss decreases,why?

jiandandan001 commented 3 years ago

Hi, how to train the model with the Adversarial Loss?
Could you give an example of the code? Thanks