Skip to content

Latest commit

 

History

History
25 lines (19 loc) · 3.87 KB

GEN_Y.md

File metadata and controls

25 lines (19 loc) · 3.87 KB

BackBone

net regressionstep psnr ssim time line
SwiftNetSlim_GFLAndMap_BN2 Backbone7x7``,在out下进行回归,decode使用bn而不是in,encoder使用InstanceNorm,GFL使用AdaptiveNorm(改动) 384P SSIMLOSS patch_loss:0.14 train_loss:0.13 psnr:30. ssim:0.88 test_loss:0.20 psnr:26.2ssim:0.83 1e5 26.2817 0.8316 18h python gfl_train_tensorboard.py --net=SwiftNetSlim_GFLAndMap_BN2 --device=cuda:1 --step=100000 --pth=SwiftNetSlim_GFLAndMap_BN2_384p_1e5_l1_ssim_IN --divisor=16 --bs=8 --l1loss --crop_size=384 --lr=0.0004 --norm --ssimloss
SwiftNetSlim2_GFLAndMap_BN2 Backbone3x3 在前一个基础 384P SSIMLOSS patch_loss:0.15train_loss:0.13 psnr:29.4 ssim:0.88 test_loss:0.21 psnr:25.8ssim:0.82 1e5 25.9860 0.8245 18h python gfl_train_tensorboard.py --net=SwiftNetSlim2_GFLAndMap_BN2 --device=cuda:0 --step=100000 --pth=SwiftNetSlim2_GFLAndMap_BN2_384p_1e5_l1_ssim_IN --divisor=16 --bs=8 --l1loss --crop_size=384 --lr=0.0004 --norm --ssimloss

GEN Y BackBone3x3 l1loss

bilinear实验结果比较好，在genY上不使用GFL

net regression step psnr ssim time line
Gen_Y_Swiftslim2_BN2Y在下采样下处理，结果使用bilinear上采样 patch_loss:0.075train_loss:0.071 psnr:29.8 ssim:0.87 test_loss:0.083 psnr:26.4ssim:0.82 2e5 26.5037 0.8232 18h python Gen_Y_train_tensorboard.py --device='cuda:0' --steps=200000 --lr=0.0004 --pth=Gen_Y_Swiftslim2_BN2_384p_2e5_l1 --divisor=16 --bs=8 --l1loss --crop_size=384 --norm --net=Gen_Y_Swiftslim2_BN2 --scale_factor=0.25
Gen_Y_Swiftslim2_BN2_ShareY结果使用上采样 & 共享encoder SPP patch_loss:0.075train_loss:0.075 psnr:29.28 ssim:0.87 test_loss:0.087 psnr:26.0ssim:0.82 2e5 26.1818 0.8220 18h python Gen_Y_Share_train_tensorboard.py --device='cuda:2' --steps=200000 --lr=0.0004 --pth=Gen_Y_Swiftslim2_BN2_Share_384p_2e5_l1 --divisor=16 --bs=8 --l1loss --crop_size=384 --norm --net=Gen_Y_Swiftslim2_BN2_Share --scale_factor=0.25
Gen_Y_Swiftslim2_BN2_SAMEY网络和主网络一样 patch_loss:0.047train_loss:0.041 psnr:29.8 ssim:0.87 test_loss:0.06 psnr:26.1ssim:0.81 2e5 26.1502 0.8163 18h python Gen_Y_train_tensorboard.py --device='cuda:1' --steps=200000 --lr=0.0004 --pth=Gen_Y_Swiftslim2_BN2_SAME_384p_2e5_l1 --divisor=16 --bs=8 --l1loss --crop_size=384 --norm --net=Gen_Y_Swiftslim2_BN2_SAME
Gen_Y_Swiftslim2_Bn2_SAME_share共享encoder SPP patch_loss:train_loss: psnr: ssim: test_loss: psnr:ssim:
Gen_Y_Swiftslim2_BN2_SAME_DownSamplegenY:在下采样分辨率下处理，结果使用GFL patch_loss:0.052train_loss:0.076 psnr:27.86 ssim:0.867 test_loss:0.079 psnr:25.49ssim:0.814 2e5 25.8214 0.8128 18h python Gen_Y_train_tensorboard.py --device='cuda:0' --steps=200000 --lr=0.0004 --pth=Gen_Y_Swiftslim2_BN2_SAME_DownSample_384p_2e5_l1 --divisor=16 --bs=8 --l1loss --crop_size=384 --norm --net=Gen_Y_Swiftslim2_BN2_SAME_DownSample --scale_factor=0.25

GEN Y: BackBone7x7 Y:l1 0: l1+ssimloss

net regression step psnr ssim time line
Gen_Y_Swiftslim_BN2很久没收敛，和训练方式有关``Y在下采样下处理，结果使用bilinear上采样 patch_loss:0.21train_loss:0.18 psnr:29.2 ssim:0.88 test_loss:0.25 psnr:26.0ssim:0.82 2e5 26.1097 0.8268 18h python Gen_Y_train_tensorboard.py --device='cuda:0' --steps=200000 --lr=0.0004 --pth=Gen_Y_Swiftslim_BN2_384p_2e5_l1_ssim --divisor=16 --bs=8 --l1loss --crop_size=384 --norm --net=Gen_Y_Swiftslim_BN2 --scale_factor=0.25 --ssimloss
Gen_Y_Swiftslim_BN2_Share很久没收敛，和训练方式有关``Y结果使用上采样 & 共享encoder SPP patch_loss:0.20train_loss:0.17 psnr:30.0 ssim:0.89 test_loss:0.24 psnr:26.2ssim:0.83 2e5 26.3842 0.8341 18h python Gen_Y_Share_train_tensorboard.py --device='cuda:2' --steps=200000 --lr=0.0004 --pth=Gen_Y_Swiftslim_BN2_Share_384p_2e5_l1 --divisor=16 --bs=8 --l1loss --crop_size=384 --norm --net=Gen_Y_Swiftslim_BN2_Share --scale_factor=0.25 --ssimloss