Invertible-Image-Rescaling icon indicating copy to clipboard operation
Invertible-Image-Rescaling copied to clipboard

Published 20 hours ago • verified publisher icon pkuxmq

Training Details

Open wenbinTse opened this issue 4 years ago • 5 comments

Hi, thank for your amazing work. I want to reproduce your work using the config train_IRN_x2.yml, but the training is slow and will cost 4 days using the default config. Do I need to notice any detail? BTW, does the default config in yml file is the correct one to reproduce the paper results?

wenbinTse avatar Jul 24 '20 12:07 wenbinTse

Yes it will cost these time to train 500k iterations. The default config is the correct one.

pkuxmq avatar Jul 27 '20 15:07 pkuxmq

Can you reproduce the DIV2k_val psnr? I train for 3 days and I think I cannot get the 44 psnr. image @wenbinTse @pkuxmq

toddwyl avatar Oct 23 '20 17:10 toddwyl

Can you reproduce the DIV2k_val psnr? I train for 3 days and I think I cannot get the 44 psnr. image @wenbinTse @pkuxmq

Is that the tensorboard show the PSNR and the paper show the PSNR_Y? @pkuxmq

toddwyl avatar Oct 24 '20 04:10 toddwyl

@ManWingloeng Yes, the training code only calculates RGB PSNR, and the testing code will calculate both RGB PSNR and Y PSNR. We report Y PSNR in the paper following the setting of other works. For IRN_x2, the final RGB PSNR should be around 40.8dB.

pkuxmq avatar Oct 24 '20 11:10 pkuxmq

@ManWingloeng Yes, the training code only calculates RGB PSNR, and the testing code will calculate both RGB PSNR and Y PSNR. We report Y PSNR in the paper following the setting of other works. For IRN_x2, the final RGB PSNR should be around 40.8dB.

Thanks your reply.

toddwyl avatar Oct 24 '20 13:10 toddwyl