Hi @INVOKERer. In MIMO-UNET, where pytorch version is 1.4, the FFT for MSFR loss is calculated as torch.rfft(image, signal_ndim=2, normalized=False, onesided=False)

Could you, please, tell me how do you port it to torch.fft.rfft in pytorch 1.8?

When you test SSIM ，did you use the function from skimage.metrics import structural_similarity？And set structural_similarity(p_numpy, label_numpy, data_range=1，multichannel=True)

Thanks for the good sharing. I concern about the ablation study on FFC (in Table 7). Did you directly replace the FT branch with FFC[41], which uses the BN operation (harmful to deblurring) ?

The gopro dataset provides two kinds of blurry images, i.e., 'blur' & 'blur_gamma'. And the original paper use 'blur_gamma' for experiments. But in your paper, it seems that you use the 'blur' dataset. I wonder if you have noticed this issue? And are all the competing methods trained on 'blur'?

In MPRNet, the Channel Attention Block (CAB) has the following structure:

`class CAB(nn.Module): def init(self, n_feat, kernel_size, reduction, bias, act): super(CAB, self).init() modules_body = []

    # ResBlock
    modules_body.append(conv(n_feat, n_feat, kernel_size, bias=bias))
    modules_body.append(act)
    modules_body.append(conv(n_feat, n_feat, kernel_size, bias=bias))
    ###

    self.CA = CALayer(n_feat, reduction, bias=bias)
    self.body = nn.Sequential(*modules_body)

def forward(self, x):
    res = self.body(x)
    res = self.CA(res)
    res += x
    return res`

In the paper, for Table 1, it is stated that ResBlock is replaced by "Res FFT-Conv Block". I wonder how do you integrate the "Res FFT-Conv Block" into CAB in your experiments? Do you just replace the ResBlock structure by "Res FFT-Conv Block" (please see my comments in the code) or you replace the whole CAB by "Res FFT-Conv Block"?

Thanks in advance.

Thanks for your wonderful deblurring work. I noticed that the provided loss curve in paper is only 120, and the epoch settting is 3000. Could you provide a complete training loss curve and complete performance curve on GoPro test dataset? Thank you.

UserWarning: Casting complex values to real discards the imaginary part (Triggered internally at ../aten/src/ATen/native/Copy.cpp:239.) Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass 这是咋回事

@INVOKERer 您好 作者 ，您的作品非常具有创造性，以至于我不得不加入我的网络尝试一下效果，我在训练的过程中遇到了相关的问题，您有遇到过如下的代码吗？

  0%|          | 0/10 [44:48<?, ?it/s]
Traceback (most recent call last):
  File "E:\UIDGain++\train.py", line 101, in <module>
    main()
  File "E:\UIDGain++\train.py", line 59, in main
    model.update_D(images_a, images_b)
  File "E:\UIDGain++\model.py", line 180, in update_D
    self.forward()
  File "E:\UIDGain++\model.py", line 128, in forward
    self.z_content_i, self.z_content_b = self.enc_c.forward(self.real_I_encoded, self.real_B_encoded)
  File "E:\UIDGain++\networks.py", line 135, in forward
    outputA = self.convA(xa)
  File "E:\anaconda3\envs\pytorch\lib\site-packages\torch\nn\modules\module.py", line 1130, in _call_impl
    return forward_call(*input, **kwargs)
  File "E:\anaconda3\envs\pytorch\lib\site-packages\torch\nn\modules\container.py", line 139, in forward
    input = module(input)
  File "E:\anaconda3\envs\pytorch\lib\site-packages\torch\nn\modules\module.py", line 1130, in _call_impl
    return forward_call(*input, **kwargs)
  File "E:\UIDGain++\networks.py", line 798, in forward
    y = torch.fft.rfft2(x, norm=self.norm)
RuntimeError: cuFFT error: CUFFT_INVALID_SIZE

Process finished with exit code 1

Thank you for your awesome code!

I am hoping you might open-source the log files you have from training. Maybe the training and validation loss as a function of epoch (and/or batch) with an estimate of the runtime?