search

xinntao / ESRGAN

ECCV18 Workshops - Enhanced SRGAN. Champion PIRM Challenge on Perceptual Super-Resolution. The training codes are in BasicSR.
https://github.com/xinntao/BasicSR
Apache License 2.0
5.91k stars 1.05k forks source link

issue with esrgan implementation #64

Open vibss2397 opened 5 years ago

vibss2397 commented 5 years ago

hi, i was making my own implementation of esrgan, except i made 2 changes, the number of rrdb blocks is 8 and i haven't included L1 loss in the perceptual loss function. The data is cell microscopy data and after 100 epochs i have been still getting output like the ones attatched in images( the generated is the 3rd image, ground truth is 4th, 1st and 2nd are bicubic and nearest neighbor). does my network need more epochs to train or is there something wrong.

P.S. i am asking this because these color losses in a normal srgan start to disappear after 10-20 epochs only.

generated_image_105 generated_image_90

HarshPathakhp commented 4 years ago

Hey, I am facing something similar too, did you find a solution to your problem?

vibss2397 commented 4 years ago

hey @HarshPathakhp, I'm sorry it didn't work i had to switch my strategy

XiaotianM commented 4 years ago

L1 loss should not deprecated in loss function