NAFNet

NAFNet copied to clipboard

hello, I have trained the nanfet-width32 on SIDD, the option file was the same as yours. The different parts are as follows.

• GPU: 8 x TitanXP
• I use accelerate (HuggingFace's tool) for training, and the training code is not BasicSR, it is a simple training template.

The training result is lower than yours. The training logger is as follows.

INFO 2023-08-07 13:35:45,530 >> start training
INFO 2023-08-07 16:20:50,954 >> epoch : 20, cur_lr :  0.000976 
                                psnr :  39.428297    ssim :  0.917009    
                                best psnr :  39.428297 @ iter 20000
INFO 2023-08-07 19:06:11,620 >> epoch : 41, cur_lr :  0.000905 
                                psnr :  39.637939    ssim :  0.918373    
                                best psnr :  39.637939 @ iter 40000
INFO 2023-08-07 21:51:42,688 >> epoch : 62, cur_lr :  0.000794 
                                psnr :  39.692169    ssim :  0.918762    
                                best psnr :  39.692169 @ iter 60000
INFO 2023-08-08 00:37:14,007 >> epoch : 83, cur_lr :  0.000655 
                                psnr :  39.771155    ssim :  0.919600    
                                best psnr :  39.771155 @ iter 80000
INFO 2023-08-08 03:22:47,570 >> epoch : 104, cur_lr :  0.000500 
                                psnr :  39.805509    ssim :  0.920232    
                                best psnr :  39.805509 @ iter 100000
INFO 2023-08-08 06:08:12,699 >> epoch : 125, cur_lr :  0.000346 
                                psnr :  39.840628    ssim :  0.920476    
                                best psnr :  39.840628 @ iter 120000
INFO 2023-08-08 08:54:00,717 >> epoch : 146, cur_lr :  0.000206 
                                psnr :  39.872308    ssim :  0.920661    
                                best psnr :  39.872308 @ iter 140000
INFO 2023-08-08 11:39:56,087 >> epoch : 167, cur_lr :  0.000096 
                                psnr :  39.877354    ssim :  0.920850    
                                best psnr :  39.877354 @ iter 160000
INFO 2023-08-08 14:25:17,954 >> epoch : 188, cur_lr :  0.000025 
                                psnr :  39.884595    ssim :  0.920922    
                                best psnr :  39.884595 @ iter 180000
INFO 2023-08-08 17:10:47,287 >> epoch : 208, cur_lr :  0.000000 
                                psnr :  39.886205    ssim :  0.920947    
                                best psnr :  39.886205 @ iter 200000
INFO 2023-08-08 17:10:54,934 >> end training

Sorry for the interruption, but I'd like to digress for a moment. Could you tell me the version of BasicSR？

Sorry for the interruption, but I'd like to digress for a moment. Could you tell me the version of BasicSR？ the version is the same as NAFNet