search

jmiller656 / EDSR-Tensorflow

Tensorflow implementation of Enhanced Deep Residual Networks for Single Image Super-Resolution
MIT License
330 stars 107 forks source link

Strange Net output PSNR? #26

Open cscyangj opened 6 years ago

cscyangj commented 6 years ago

Why this happen?

I did not change net params, and using dataset of my own. PSNR is normal at first , while a big degrade then happed. psnr

hrahoy commented 6 years ago

In my case, the lower learning rate solved this problem.

purse1996 commented 6 years ago

After upsamle by sub-pix, there is an annotated code here output = slim.conv2d(x_reslut,3,[3,3])， it seems to a 3x3 convlution after upsample. When I use the convolution, the same problem happen as you. When removing it, it gets normal.

ZhenyF commented 6 years ago

@purse1996 I meet the same problem as you: when I completely implement the upscale method in the paper (add a convolution after the upsampling layer), the output is lower than 20 but when I remove the final conv layer is goes fine. I also tried the code using pytorch. It turns out that the paper auther's method on pytorch works pretty good. Do you know why is that?

purse1996 commented 5 years ago

In fact, I don't solve the problem and just ignore the convolution layer. If you have any idea about it, please tell me. Thank you very much. Additionally, Do you meet the issue that the gpu-util is always zeros when using the code, Do you have any solutions? Thanks a lot.

ZhenyF notifications@github.com 于2018年8月26日周日 上午4:19写道：

@purse1996 https://github.com/purse1996 I meet the same problem as you: when I completely implement the upscale method in the paper (add a convolution after the upsampling layer), the output is lower than 20 but when I remove the final conv layer is goes fine. I also tried the code using pytorch. It turns out that the paper auther's method on pytorch works pretty good. Do you know why is that?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/jmiller656/EDSR-Tensorflow/issues/26#issuecomment-415994066, or mute the thread https://github.com/notifications/unsubscribe-auth/AlZDSvCoixTbIbdYvMx9BISj-Joo83_vks5uUbFXgaJpZM4SxOFw .

ZhenyF commented 5 years ago

@purse1996 One of the reasons causes your low GPU util might be your data processing period in CPU is long, so your GPU need to wait for data from CPU.

suke27 commented 5 years ago

seems psnr is poor during training, always less than 15. I did not add output = slim.conv2d(x_reslut,3,[3,3])， after upsample, anybody know what is the reason