Skip to content

Latest commit

 

History

History
60 lines (46 loc) · 8.47 KB

ex.md

File metadata and controls

60 lines (46 loc) · 8.47 KB

LOL

对于数据集的指标,使用没有光照scalar的模型做infer。和其他人的一样。 ulimit -n 2048 nohup python >log.out

过拟合1:patch与total_image关于image_size的过拟合 过拟合2:train/test的过拟合 gen_y_unet_160p_2e5_l1.pth step : 88000 max_psnr : 20.69818210016854(+1) max_ssim : 0.8325258769386841(+0.03)

net step psnr ssim time line
unet其结果模糊不清尽管能够做到denoise 1e5 19.6919 0.8089 4h python train.py --net='unet' --step=100000 --pth=unet_160p_1e5_l1 --divisor=16 --bs=8 --l1loss --crop_size=160
unet64 1e5 20.1542 0.7777 4h python train_tensorboard.py --net='unet64' --step=100000 --device=cuda:0 --pth=unet64_160p_1e5_l1 --divisor=16 --bs=8 --l1loss --crop_size=160
FullConv_SwiftNet 1e5 0.7805 19.5069 5h python train_tensorboard.py --net='FullConv_SwiftNet' --device=cuda:0 --step=100000 --pth=FullConv_SwiftNet_160p_1e5_l1 --divisor=32 --bs=8 --l1loss --crop_size=160 --lr=0.0004
hdr1没有训练patch和整张图的过拟合(因为子分辨率变换算子与原分辨率算子相似),但是有train/test的过拟合,grid_sample采用的是bilinear(x,y),最关键的guided-map(z)没有采样! 1e5 25.6277结果较清晰 0.7534但存在很大的噪声 19h python train_tensorboard.py --net='hdr1' --step=100000 --device=cuda:0 --pth=hdr1_384p_1e5_l1 --divisor=1 --bs=8 --l1loss --crop_size=384
Res18Net1 dims=[16,24,32,48,64]不能收敛 dims=[64,64,64,64,64]收敛更差 1e5 10h python train_tensorboard.py --net=Res18Net1 --step=100000 --device=cuda:0 --pth=Res18Net1_384p_1e5_l1 --divisor=16 --bs=8 --l1loss --crop_size=384

Swiftnet +guided filters

收敛标准在0.02左右 InstanceNorm能够有效解决过拟合1 因为使用了Batch norm并不能够有效解决过拟合2

net step psnr ssim time line
swiftnetoutput是1/4 上采样 结果模糊的要死 1e5 19.2726 0.7124 8h python train_tensorboard.py --net='swiftnet' --device=cuda:0 --step=100000 --pth=swiftnet_160p_1e5_l1 --divisor=32 --bs=8 --l1loss --crop_size=160 --lr=0.0004
SwiftNet_GuidedFilteroutput使用GuidedFilter,回归x4,loss都是在x4下,psnr和ssim在原分辨率下回归到0.01非常好,过拟合1:train_loss:0.05很差,过拟合2也存在但较轻eval_loss:0.06 1e5 22.3718 0.7298 18h python gf_train_tensorboard.py --net=SwiftNet_GuidedFilter --device=cuda:1 --step=100000 --pth=SwiftNet_GuidedFilter_384p_1e5_l1 --divisor=32 --bs=8 --l1loss --crop_size=384 --lr=0.0004
swiftnetslim 特征变为[16,32,64,64]原来[64,128,256,512] 约等于w=0.2也是能回归的0.02 2e5 17.7897 0.6939 3h python train_tensorboard.py --net='swiftnetslim' --device=cuda:0 --step=200000 --pth=swiftnetslim_160p_2e5_l1 --divisor=32 --bs=8 --l1loss --crop_size=160 --lr=0.0004
SwiftNetSlim_GuidedFilterLayerAndMap在out下进行回归,swiftslim没有使用norm,GFL使用AdaptiveNorm``patch_loss:0.024 train_loss:0.056(过拟合1) psnr:23.7 ssim:0.86 test_loss:0.067 psnr:22.2(-1.5)ssim:0.81(-0.05) 1e5 22.9125 0.8107 18h python gfl_train_tensorboard.py --net=SwiftNetSlim_GuidedFilterLayerAndMap --device=cuda:0 --step=100000 --pth=SwiftNetSlim_GuidedFilterLayerAndMap_384p_1e5_l1 --divisor=32 --bs=8 --l1loss --crop_size=384 --lr=0.0004
SwiftNetSlim_GuidedFilterLayerAndMap在out下进行回归,swiftslim使用InstanceNorm,GFL使用AdaptiveNorm``patch_loss:0.023 train_loss:0.023(没有过拟合1) psnr:29.57 ssim:0.872 test_loss:0.40psnr:25.8(-4过拟合2严重)ssim:0.815(-0.06严重) 1e5 25.8456 0.8169 18h python gfl_train_tensorboard.py --net=SwiftNetSlim_GuidedFilterLayerAndMap --device=cuda:1 --step=100000 --pth=SwiftNetSlim_GuidedFilterLayerAndMap_384p_1e5_l1_IN --divisor=32 --bs=8 --l1loss --crop_size=384 --lr=0.0004 --norm
SwiftNetSlim_GFLAndMap_BN ,在out下进行回归,decode使用bn而不是in,encoder使用InstanceNorm,GFL使用AdaptiveNorm``patch_loss:0.02-0.024 train_loss:0.024 psnr:29.7 ssim:0.87 test_loss:0.037(过拟合2)psnr:26.4(-3)ssim:0.819(-0.05) 1e5 26.4670 0.8207 18h python gfl_train_tensorboard.py --net=SwiftNetSlim_GFLAndMap_BN --device=cuda:0 --step=100000 --pth=SwiftNetSlim_GFLAndMap_BN_384p_1e5_l1_IN --divisor=32 --bs=8 --l1loss --crop_size=384 --lr=0.0004 --norm
SwiftNetSlim_GFLAndMap_BN 在out下进行回归,decode使用bn而不是in,encoder使用InstanceNorm,GFL使用AdaptiveNorm SSIMLOSS``patch_loss:0.15 train_loss:0.13 psnr:30.0 ssim:0.88 test_loss:0.20 psnr:26(-4)ssim:0.83(-0.05) 1e5 26.0504 0.8297 18h python gfl_train_tensorboard.py --net=SwiftNetSlim_GFLAndMap_BN --device=cuda:0 --step=100000 --pth=SwiftNetSlim_GFLAndMap_BN_384p_1e5_l1_ssim_IN --divisor=16 --bs=8 --l1loss --crop_size=384 --lr=0.0004 --norm --ssimloss
SwiftNetSlim_GFL_SN 全部替换为SwitchableNorm保留adaptiveNorm形式,SSIMLOSS``patch_loss:0.14 train_loss:0.12 psnr:30.3 ssim:0.89 test_loss:0.20 psnr:26.3(-4)ssim:0.83(-0.06) 1e5 26.4075 0.8298 18h python gfl_train_tensorboard.py --net=SwiftNetSlim_GFL_SN --device=cuda:1 --step=100000 --pth=SwiftNetSlim_GFL_SN_384p_1e5_l1_ssim --divisor=16 --bs=8 --l1loss --crop_size=384 --lr=0.0004 --norm --ssimloss
SwiftNetSlim_GFLAndMap_BN2 ,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) 256P SSIMLOSS patch_loss:0.13-0.15 train_loss:0.15 psnr:27.15 ssim:0.877 test_loss:0.212 psnr:25.36(-2)ssim:0.82(-0.05) 1e5 25.6620 0.8249 18h python gfl_train_tensorboard.py --net=SwiftNetSlim_GFLAndMap_BN2 --device=cuda:1 --step=100000 --pth=SwiftNetSlim_GFLAndMap_BN2_256p_1e5_l1_ssim_IN --divisor=1 --bs=8 --l1loss --crop_size=256 --lr=0.0004 --norm --ssimloss
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 在前一个基础上conv改为3x3而不是7x7 256P SSIMLOSS patch_loss: 0.15 train_loss:0.145 psnr:28.47 ssim:0.875 test_loss:0.21 psnr:25.6(-3)ssim:0.822(-0.05) 1e5 25.7234 0.8224 18h python gfl_train_tensorboard.py --net=SwiftNetSlim2_GFLAndMap_BN2 --device=cuda:1 --step=100000 --pth=SwiftNetSlim2_GFLAndMap_BN2_256p_1e5_l1_ssim_IN --divisor=16 --bs=8 --l1loss --crop_size=256 --lr=0.0004 --norm --ssimloss
SwiftNetSlim2_GFLAndMap_BN2 Backbone 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

Guided Filter Network

收敛标准在0.02左右 因为使用了norm所以有更好解决过拟合2

net step psnr ssim time line
DeepGuidedFilter 回归到0.05左右,过拟合1:train_loss在0.07,过拟合2:存在但不稳定,有时候轻微有时候严重 1e5 21.8926 0.7401 20h python train_tensorboard.py --net=DeepGuidedFilter --step=100000 --device=cuda:0 --pth=DeepGuidedFilter_384p_1e5_l1_025 --divisor=1 --bs=8 --l1loss --crop_size=384 --scale_factor=0.25
DeepGuidedFilterAndMap 回归到0.05左右,过拟合1效果同上,但过拟合2交情较轻 1e5 21.9899 0.7865去躁了所以好了点 20h python train_tensorboard.py --net=DeepGuidedFilterAndMap --step=100000 --device=cuda:1 --pth=DeepGuidedFilterAndMap_384p_1e5_l1_025 --divisor=1 --bs=8 --l1loss --crop_size=384 --scale_factor=0.25
DeepGuidedFilterLayer回归到0.05左右,过拟合1:train_loss在0.07上,过拟合2:几乎没有甚至eval比train在loss上更低 1e5 22.1607 0.7719 20h python train_tensorboard.py --net=DeepGuidedFilterLayer --step=100000 --device=cuda:0 --pth=DeepGuidedFilterLayer_384p_1e5_l1_025 --divisor=1 --bs=8 --l1loss --crop_size=384 --scale_factor=0.25
DeepGuidedFilterLayerAndMap回归到0.05左右过拟合1:train_loss在0.07上,过拟合2:几乎没有甚至eval比train在loss上更低 1e5 21.9514 0.7808去躁了所以好了点 20h python train_tensorboard.py --net=DeepGuidedFilterLayerAndMap --step=100000 --device=cuda:1 --pth=DeepGuidedFilterLayerAndMap_384p_1e5_l1_025 --divisor=1 --bs=8 --l1loss --crop_size=384 --scale_factor=0.25