Thank you very much for sharing, I have some questions here.Is the input of DIQA also patch? Can the whole picture be input directly? I did not see the details of patch input in the paper.

Hello, I read your paper and run your code in my machine, I found the mathod defined in IQADataset.py as following:

def errormap_process(patch, patch_ref):
p = 0.2
errormap = np.abs(patch_ref.numpy(), patch.numpy())
errormap = np.power(errormap, p)
errormap = torch.from_numpy(errormap)
errormap = errormap.unsqueeze(0)
errormap = F.interpolate(errormap, size=(28, 28))
errormap = errormap.squeeze(0)
return errormap

and in your paper, it said that error map is defined used the exponent difference function, so I think the bold code of upper method should be:
errormap = np.abs(patch_ref.numpy() - patch.numpy())

I believe it was due to your hand mistake.

请问在iqadataset.py里的第27和29行有什么具体意义嘛。。。我知道做了什么，但是感觉没什么用啊

thanks for your code. I train it but loss is nan, error shows that the dividend is 0. I don't know the reason. Can you tell me answer?

When I run train_ step1.py, it said, "typeerror: can't convert CUDA tenant to numpy. Use Tensor.cpu () to copy the tensor to host memory first.”
Can you tell me the version of torch package? I think the problem may be the package version.
thank u.