search

Lornatang / SRGAN-PyTorch

A simple and complete implementation of super-resolution paper.
Apache License 2.0
419 stars 105 forks source link

loss of generator #69

Closed Jihnjockey closed 1 year ago

Jihnjockey commented 1 year ago

image Why do you plus the pixel loss? I remembered that the author didn't add that loss part.

Lingyun-Wang commented 1 year ago

image Why do you plus the pixel loss? I remembered that the author didn't add that loss part.

I agree with you. In that paper, there are three kinds of SRGANs which are investigated in section 3.4: SRGAN-MSE (pixel-loss and adversarial-loss), SRGAN-VGG22(content-loss and adversarial-loss), and SRGAN-VGG54(content-loss and adversarial-loss).

Lornatang commented 1 year ago

For pixel loss， I research w/o pixle loss and w/ pixel loss, if you not increase pixel loss, the problem is abnormal color, difficult to converge... But I can't find the paper reproduced code by the original author, so at present, it is implemented according to the best loss method~~

Lingyun-Wang commented 1 year ago

For pixel loss， I research w/o pixle loss and w/ pixel loss, if you not increase pixel loss, the problem is abnormal color, difficult to converge... But I can't find the paper reproduced code by the original author, so at present, it is implemented according to the best loss method~~

Thanks for your reply.