about SRResNet

[denghuiru]

Hi, happy to see that you are trying to implement SRGAN using caffe, I am wonder whether you have implemented the PSNR of SRResNet in paper proposed.

[ShenghaiRong]

You can see the benchmarks in readme.md . The PSNR of SRResNet-MSE in paper proposed is 32.05, but right now my best result is 31.26. I'm still trying to modify the params and finetuning the networks.

[denghuiru]

@ShenghaiRong I have trying to add sub-pixel layer as you said, however, after modify the reshape_layer.cpp && .hpp, it still can not sucessfully make caffe, I am wonder if you have also modified other codes. Would you paste it into your code?

[ShenghaiRong]

Did you modify the caffe.proto? Add optional int32 pixelshuffler = 4 [default = 1]; to your reshape_param in caffe.proto. You can refference my caffe.proto .

[denghuiru]

@ShenghaiRong Hi, I am wonder have you trained the SRGAN-MSE networks, and have you got some good results?

[ShenghaiRong]

Yes，I have trained them. But the results were not good.The results showed that the impacts of GAN were not obvious.

[denghuiru]

@ShenghaiRong The SRGAN-MSE is not very obviously if you test the results on Set5 just as the paper said, Have you test the corresponding PSNR values,dose it match with the paper proposed?

[ShenghaiRong]

Have you trained the SRGAN-MSE? My PSNR of the SRGAN-MSE is 31.24 and the SSIM is 0.5188.

[denghuiru]

@ShenghaiRong yeah, I have also try to implement the SRGAN-MSE, but ,however, the PSNR dose not decrease so quickly, I think, because the SRGAN-MSE results are not so obviously, so if we can reach the psnr proposed in the paper,that will demonstrate the this training is correctly.

[Jee-King]

Hello,I have modified the caffe.proto, add optional int32 pixelshuffler = 4 [default = 1].but it have a error: Error parsing text-format caffe.NetParameter: 702:20: Message type "caffe.ReshapeParameter" has no field named "pixelshuffler". How to solve it?

[denghuiru]

@ShenghaiRong hi, i have saw that you upload the segan_train.py, does this is the new code to train the srgan? And have you verify the correctness of this code?

[ShenghaiRong]

Right now I have uploaded the file of pycaffe_srgan. I'm still training the new version of srgan. But I'm busy with my dissertation. Maybe you can help me verify the correctness of this code :)

[pangyunping]

@Ji-qing You should edit your caffe.proto,add
optional int32 pixelshuffler = 4 [default = 1]; to your function ReshapeParameter {

[Imanpour]

Hi,

Just for adding pixel shuffling layer in caffe what shold we do? I mean the exact changes in Caffe files for using this layer.

Thanks

[ghost]

Thanks for your answer. I did that. Does this work for any up scaling factor?

On Dec 26, 2017 07:36, "rsh" notifications@github.com wrote:

@Imanpour https://github.com/imanpour you can see the Implementation Details 1：replace your reshape.hpp , cpp.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/ShenghaiRong/caffe_srgan/issues/1#issuecomment-353926146, or mute the thread https://github.com/notifications/unsubscribe-auth/AYg65li8880yGz8wuE21pz0XtBE0kOanks5tEJPzgaJpZM4MwX4I .

[ShenghaiRong]

Maybe ... :)

[ShenghaiRong]

@Imanpour you can see the Implementation Details 1 in the readme.md ：replace your reshape.hpp , cpp.