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
6.02k stars 1.07k forks source link

Changing `upscale` factor? #24

Open MasterScrat opened 5 years ago

MasterScrat commented 5 years ago

Is there any easy way to tune the upscale=4 factor? would that need a specifically trained network?

xinntao commented 5 years ago

Hi @MasterScrat Yes, need to train/fine-tune specifical networks for other factors like x2, x3.

Maybe we can first upsample by 4, then use some easy cubic or bilinear downsampling methods, but I am not sure the final results will be OK.

MasterScrat commented 5 years ago

I see. How much time and hardware resources were necessary to train RRDB_PSNR_x4?

xinntao commented 5 years ago

@MasterScrat I used a Titan XP and it cost about one week. really slow...

You may speed up by multiple-GPU training.

DeltaDesignRus commented 5 years ago

@xinntao Is it possible to use scale x1? I want to train the network to remove artifacts of compression jpg without scale.

xinntao commented 5 years ago

@DeltaDesignRus Though it is possible to use scale x1 (remove the upsampling layers), it is recommended to use models that are designed for compression artifact removal. There may be some better designs for the models of compression artifact removal, like U-Net for speeding up.