search

megvii-research / NAFNet

The state-of-the-art image restoration model without nonlinear activation functions.
Other
2.08k stars 256 forks source link

training nafnet_sidd_width_32 without BasicSR #118

Open HLImg opened 11 months ago

HLImg commented 11 months ago

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

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
Amy-214 commented 10 months ago

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

HLImg commented 10 months ago

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