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

Training ESRGAN from scratch without VGG feature extractor #81

Open donCorleone97 opened 4 years ago

donCorleone97 commented 4 years ago

Hi, inspired from your model, I have written a similar ESRGAN model with spectral_normalization and a few other tweaks. Although mostly, the underlying model architecture is same. I am also using VGG19_54 for feature extraction. I'm trying to train the model to super resolve images which VGG19 has never been trained on.

My generator is not producing the highest quality images, there are a few distortions, although the perception of image is excellent. I believe the fault lies in the discriminator not being able to discriminate between fake and real images. Will you please suggest something I can do to mitigate it?

I was thinking if I have to remove VGG_19 part, I can extend the discriminator to be a proper robust CNN OR I can train a VGG19 model without pretrained weights on my own dataset and then use it to extract features. Will that work?

kpagels commented 4 years ago

@donCorleone97 did you figure out how to solve this?