the model was trained on DIV2k? test SET5 on it , but the psnrs are 17~30.

[ChChwang]

the model was trained on DIV2k? test SET5 on it , but the psnrs are 17~30。 Can it be reach the same result of the paper ?

[ChChwang]

add --matlab_bicubic=True and the PSNR on Set5 is 31

[kong1060553768]

@ChChwang Why have I set'- matlab_bicubic = True', step number: 1774710, scale: 3, but PSNR on set5 is 31.5

[ChChwang]

@ChChwang Why have I set'- matlab_bicubic = True', step number: 1774710, scale: 3, but PSNR on set5 is 31.5

me too, I can not get the result on the paper

[kong1060553768]

Are you the author's checkpoint? What is the average PSNR you get on set5?

[ChChwang]

Are you the author's checkpoint? What is the average PSNR you get on set5?

not the author's checkpoint, I trained it by myself, PSNR=31, about 3% lowers the paper

[kong1060553768]

@ChChwang为什么我设置'-matlab_bicubic = True'，步号：1774710，比例：3，但是set5上的PSNR是31.5

我也是，我无法在纸上得到结果

你是作者的检查站吗？你在set5上获得的平均PSNR是多少？

不是作者的检查点，我自己训练，PSNR = 31，大约3％降低了论文

How many steps have you trained? But in scale = 3, PSNR = 31, the image of PSNR and bicubic effect is similar, do you find a solution?

[Missdonghui]

Why does the --matlab_bicubic increase the value of PNSR

[kong1060553768]

@Missdonghui Because the checkpoint given is that the author is trained by the data sampled under matlab.

[Missdonghui]

@Missdonghui Because the checkpoint given is that the author is trained by the data sampled under matlab.

Thankyou a lot

[kong1060553768]

@Missdonghui Can you tell me the result of your reproduction?

[Missdonghui]

I am very sorry that I did not use it for the test, because of the computer configuration, but the PSNR without it is 19.3223

[kong1060553768]

Okay, I didn't get very good results.

[Missdonghui]

So what do you get

[kong1060553768]

Change a network

[Missdonghui]

Can you tell me more about it

[kong1060553768]

I gave up RDN and used EDSR

[Missdonghui]

Okay,Thanks

[Mmmofan]

Are you the author's checkpoint? What is the average PSNR you get on set5?

not the author's checkpoint, I trained it by myself, PSNR=31, about 3% lowers the paper

I trained on DIV2K for 30 epochs(about 27673 steps per epoch with batch size 16), test on Set14 and Set5, the PSNR is get higher from the begining, but getting lower from epoch 7...
The maximum PSNR is 36.88 on Set5 33.00 on Set14...
I don't know why I can't get the result author does...

[huang-junhong]

Hellow
For your problem i think you had set something wrong.
limit by the GPU i had test the RDN(D10C6G32) witch paper's net is (D16C8G64) use DIV2K without flap, and batchsize set 16, and rewrite code by myself.
while 200K~250K PSNR on DIV2K is 32.10 when almost 400K max is 32.12
So I think author's result not fake.
More I test D10C6G32 while train, the cruve isn't like paper and the converge speed is ver slow.
After 100K iter it's PSNR not achieve 37.2, then I shut down it.
And I try D10C8G64 , the curve like paere whith note D10C6G32. But this test just begin.