zhaominyiz / c3-stisr Goto Github PK

I made the test with your code and get the same performance as you reported in Readme. That's cool, however, when I visualize the reconstruct result with visualize option in your code, I can find it has some unusual visual artifact.

restored picture is as followed, and as you can see, restoration picture shows decreased visual quality and background color is getting a little different.

Performance is exactly same as you reported, so I ask you whether this artifact a usual one.

Best

I have fused the code into TPGSR, but after 500 epochs of training, the result is about 5% lower than the result in the article.
(The accuracy obtained by loading the parameter file you posted is normal.)

hey, it seems if we want to run the test/train using your arch, we need to modify the base.py from interface directory (especially the 293rd line about generator_init.
can you share the modified base.py?

Hi, author!
When I want to load the pretrained model as you have advised in #7 (comment), however, I met problem in the line params = torch.load('ckpt/BCN_correct_model.pt') . I couldn't find this pretrained model BCN_correct_model.pt in your release. I wonder if it is convenient for you to share this? Thank you !

I don't have any idea what this file contains

I have given my start!
I also met the same issue, could you please provide me with the full code?
My email is [email protected].
Thanks a lot!

Hi~ I have given my start!
I also met the issue of running the training code, could you please provide me with the full code?
My email is [email protected].
Thank you very much!

如果我只想做单张图像的超分辨率结果，请问我应该怎么改代码？

image_lr = transforms.ToTensor()(Image.open('./val_data/20230317.jpeg').convert('RGB')).unsqueeze(0).cuda()
print(image_lr.shape)
images_sr = model(image_lr) 

但这样会报错，我用的crnn.pth

conv_feature = torch.cat([conv_feature, position2d],1) # batch, 128(64+64), 32, 128
RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 240000 but got size 1024 for tensor number 1 in the list.

Hi, have a nice day! Thanks for your impressive work. Recently I was trying to reproduce your code. However, I have encountered a problem regarding image quality, i.e. PSNR and SSIM. The avg PSNR reported in the paper is 21.51 dB, but the log below is 20.38 dB. and the same problem is encountered with SSIM. I don't know why this is the case. I would be grateful if you could reply :D

evaling easy
[2022-09-05 14:25:13]	PSNR 22.13 | SSIM 0.8532	
save display images
sr_accuray_iter0: 65.60%
lr_accuray: 37.49%
hr_accuray: 76.41%
best_easy = 65.60%*
evaling medium
[2022-09-05 14:25:29]	PSNR 18.98 | SSIM 0.6449	
save display images
sr_accuray_iter0: 53.44%
lr_accuray: 21.40%
hr_accuray: 75.05%
best_medium = 54.43%
evaling hard
[2022-09-05 14:25:45]	PSNR 19.75 | SSIM 0.7125	
save display images
sr_accuray_iter0: 39.91%
lr_accuray: 21.15%
hr_accuray: 64.56%
best_hard = 39.91%*
saving best model

您好，通过您的文章，我对您的工作很感兴趣，请问大概什么时候公布代码呢