This error maybe due to breaking changes in torchvision module.

The following line throws an error:

from torchvision.models.utils import load_state_dict_from_url

The following change works:

from torch.hub import load_state_dict_from_url

line 122 should be like
@classmethod
def has_split(cls, split):
return split in cls().get_splits()

Hi
I am unable to find the paper for ninasr. Is it possible to provide the paper or a link ? Thanks

I have also a very intersting question to ask you that when I test the model on div2k validation set, my own codes always tells me that my cuda memory is run out, but in your code the same model works perfectly on the div2k validation set so how do you make this work?

hi

sorry for bothering, and recently i have used your scripts which really great work.

may i ask one question that "i have found "RandomCrop" is appended to training dataset, and i wonder how it works?

thanks a lot for your attention and support.

br,
sc

Hi there!

First of all, amazing work! I am developing a CV library called glasses and I was thinking to also add SR models and I came across your repo.

What if we merge the two? Would you like to talk about it?

Let me know!

Thank you :)

Cheers,

Francesco

Alot of these networks are still too heavy for cpu inference would you be open to a PR to add these models ?

First, thanks a lot for creating this one-stop repository!

I want to use the NinaSR in one of my works and would like to cite it. Any chance you could share that information on the readme file?

If I run the sample code on image indexed 5:

from torchsr.datasets import Div2K
from torchsr.models import ninasr_b0
from torchvision.transforms.functional import to_pil_image, to_tensor

# Div2K dataset
dataset = Div2K(root="./data", scale=2, download=False)

# Get the first image in the dataset (High-Res and Low-Res)
hr, lr = dataset[5]

# Download a pretrained NinaSR model
model = ninasr_b0(scale=2, pretrained=True)

# Run the Super-Resolution model
lr_t = to_tensor(lr).unsqueeze(0)
sr_t = model(lr_t)
sr = to_pil_image(sr_t.squeeze(0))
sr.show()

I get the results dotted with artefact pixels. See blue, red dots:

My env:

certifi==2023.7.22
charset-normalizer==3.2.0
filelock==3.12.2
fire==0.5.0
idna==3.4
Jinja2==3.1.2
MarkupSafe==2.1.3
mpmath==1.3.0
networkx==3.1
numpy @ file:///D:/bld/numpy_1688887186903/work
Pillow==10.0.0
requests==2.31.0
six==1.16.0
sympy==1.12
termcolor==2.3.0
torch==1.8.1
torchsr==1.0.4
torchvision==0.9.1
typing_extensions @ file:///home/conda/feedstock_root/build_artifacts/typing_extensions_1688315532570/work
urllib3==2.0.4

It feels like so some sort of overflow but I could not find it.

Hello, I really love your program because it helps me a lot of understanding the image super-resolution via pytorch. However, I want to tell you something which I have found during my usage of running your code. I have compared the value of a same image group (Sr images and Hr images) from your code and other official methods, and I found a little difference when it executes the calculation of SSIM and PSNR between the PIQ (yours) and the official method of matlab (from the code of RCAN). Then I found the way of calculation of SSIM and PSNR from RCAN and SwinIR has the same value which is not the same as the PIQ method. So would you like that I provide you the calculation method from SwinIR, and we could make an intergration in to your code? If you are interested in that, don't be hesitate of contacting me.

Hi, I am really grateful that you give us a very clear implementation of these SR models, by the way I would like to ask you about the parameters(optimiser, data-augement, epochs ,etc) that you have choosen in the reproduction during the training, especially for the EDSR huge, because I have problems when I apply the parameters in the paper on your training models, as I could not reach the pnsr and ssim as high as the paper. Your response can be really helpful for me and I would like to thank you in advance.