junshk / cincgan-pytorch Goto Github PK

Thank you Junshk so much for your good work. I tried to implement the code you wrote but since the instruction is quite ambiguous so I could not run it successfully. Could you help clarify how the training data is allocated? For instance, in the paper, it is said the HR images Z are from 400-800 training images from DIV2K, and input LR X are from 1-400 images. Besides, for Y (intermediate) are from the bicubic downsampling of HR images of Z, which means Y are also numbered as 401-800. However, when I tried the code it is said that '../dataset/DIV2K/DIV2K_train_HR/0015.png', which means you did not actually divide the HR and LR raw data from DIV2K right?
Furthermore, the code is quite hard to understand, so could you please update or make a concise version? Thanks for you help.

How can we train CinCGAN with our own dataset?
Kindly help in this matter.

Hi, what is your loss after training step 1? How can I know my loss is stable to begin training step 2? Is there a rough guideline?

I used 10 GPUs and after 30 epochs now the loss is

===> Epoch28: Loss: idt 0.000000 0.037730 cyc 0.000000 0.005994 D 0.000000 0.464715, G: 0.000000 0.284303, psnr_hr: 16.187803, psnr_lr 7.150141

and it is not improving

there only 800 iamges in file DIV2K_train_LR_mild, why the program try to find 0801x4m.png in train stage?shuold I put validation data in file DIV2K_train_LR_mild(it conflicts with the descrobtion in the paper)?can anyone slove the problem?

This works is very interesting. However, I don't understand the implementations. Does the code have concise version?

Hi is the download link downloading the correct model?
The names of the downloaded model is EDSR_x4.py insteand of EDSR_baseline_x4.pt as suggested in the code.

After I renamed the file. It seems the model dimension does not match.

Traceback (most recent call last):
File "/home/shitaili/workspace/jupyter_notebooks/sr/CinCGAN-pytorch/code/model/edsr.py", line 65, in load_state_dict
own_state[name].copy_(param)
RuntimeError: The expanded size of the tensor (64) must match the existing size (256) at non-singleton dimension 0. Target sizes: [64, 3, 3, 3]. Tensor sizes: [256, 3, 3, 3]

about the result of NTIRE 2018 track 2, has somebody test the NTIRE 2018 track 2 valid dataset? I think the paper's result is not true because even the bicubic result and the BM3D+EDSR result is not true.

Could you provide the instructions to use your code for training and testing.
Thank you.

I'd like to see the results of the paper reproduction. Could you upload some pictures?
thank you so much

I m getting this error. Can anyone help?

Hi,

I'm getting all losses as nan, and I find out that in main.py, around line 200, dn_ = model0 returns all nan outputs. I print out the input_v and the input_v looks correct. Would you please let me know what shall I do?

Thanks!

The model is trained by the default option. The reconstructed img is in the "result" folder after 400 epoch trained.

problem：
when I load EDSR_x4.pt ，and use the load_state_dict method in EDSR.py，the problem shows:

GHR.load_state_dict(Loaded,True)
KeyError: 'unexpected key "body.16.body.0.weight" in state_dict'。

GHR.load_state_dict(Loaded,False) can run the code ，but the result of sr is pool，so I think whether the structure code of EDSR.py you write maybe not complete？

How extend your code to "Multiple Cycle-in-Cycle Generative AdversarialNetworks for Unsupervised Image Super-Resolution"
Best

hi, thanks for you wonderful work~
when i got through your code , finished the 1st loss print, it's get a error "CUDA out of memory"
Do you know how to fix it ?
thanks for your time ~

May I ask such a question？

I saw the code and I think it is not splitting the denoising and fine tuning step as mentioned in the original paper. It would be appreciated if anyone has any idea how to split it up.

Thanks!