hejingwenhejingwen/cresmd

(ECCV 2020) Interactive Multi-Dimension Modulation with Dynamic Controllable Residual Learning for Image Restoration

cresmd's Issues

question about "zero point " problem with the pretrained global scale weights

Hi,
I use your pretrained model 2D_modulation.pth to test 2D problem.
I load the weights and find the global scale conv weights as follows, which is corresponding to the "zero point " problem:

print(list(model.children())[0].weight)
Parameter containing:
tensor([[ 1.1378e-05, -8.9791e-05],
        [-5.6083e-05,  6.1680e-05],
        [-4.5423e-05,  3.8486e-05]], requires_grad=True)

print(list(model.children())[0].bias)
Parameter containing:
tensor([-0.0140,  0.0141,  0.0138], requires_grad=True)

If the blur & denoise 2D problem cond vector to "zero point" [0,0], and the weight will be useless, the bias dominate.
And the bias does not be 0.

So, what's the problem?
Why do you just set the global scale linear bias=False?

source code

Hi, can you share the source code? Thanks

Pretrained model

hello, I am trying to do JPEG debloking with different JPEG compression ratio, I think your work may be very helpful, would you mind providing your pretrained model(noise+JPEG, blur+JPEG and noise+blur+JPEG).

Thank you!

Will the 3D pretrained-model be released?

deblock+denoise+deblur

Also I want to reprocude the model by other (our own) dataset.
How to make the noise/blur/jpeg artifact?
What is the sequence? And what are the level range of these three processes? What kind of noise/blur should we use?

cond variable assignment

hi, jingwen, a quick question.

test.py Line57 cond variable is repeatedly divided by cond_norm, making itself rapidly decreasing to 0. Is it correct?

hi can you give me your email?

my python==3.8 ,when i run train.py,it shows the error:"TypeError: expected str, bytes or os.PathLike object, not NoneType".how to slove it? my email:[email protected].
thank u!

dose it support HD image for denoise

dose it support HD image for denoise