Hello,
I am confused about why you did not use Batch Normalization or Instance Normalization for your architecture?
I do some experiment about the Normalization layer, I add Instance Normalization layer for all encoder and decoder, where each normalization layers was put between convolution layer and ReLU function, but the performance got worse after I put the normalization layers compared to the initial DMPHN architecture, the performance would get worse started from the first epoch.

Did you test this before and why did this happen? Thank You!

谁能告诉我，为什么要分多个patch，卷积操作难道输入输出不一样么？而且多patch，你算loss时候也要平均，也没有多训练数据的意思吧？
how can tell me the reason for multi-patch, the CNN is not non-local, there is no difference for multi-patch and whole image.??? besides, it updates loss once per images(batch), so, in my understanding, multi-patch can not lead to more training data.

Does it really need such a long time for reproduction ？OMG，this is the first time I've seen so much epochs.

Repository requires scikit-image, which is not mentioned in the README

Could you give some details of your setting of VideoDeblurring testing set？
I found that I can get the same PSNR results on GOPRO dataset， but get a lower PSNR results on VideoDeblurring dataset when I run your codes as the checkpoint you provided

Hi, author, thank you for providing the code, could you tell me how long it takes to train a complete code？It takes me about two weeks to train for the model DMPHN_1_2_4 on 2080Ti. Is this normal?

Thank you for your awesome code!

Recently,i have read your paper Event-guided Multi-patch Network with Self-supervision for Non-uniform Motion Deblurring，i'm really interested in it,could you open source the code about the event version? thank you!

I run the code and tested images. Averagely it cost 200ms to test one image in size of 1280x960, which is far slowly than the paper said. I tested GTX1080TI, GTX TITAN X, GTX2080TI. How could you get so fast inference speed as your paper stated?

In Paper, the final conv layer in decoder, the out_channel mentioned is 32 but in code it is given as 3. Please clarify which is correct for the given results and analysis made in paper.

Thank You

运行1000个epoch，模型就已经开始收敛

Hello, is it in Windows environment or Ubuntus environment

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?

Hello sir,
I have interesting image deblurring and i searched your code.

I checked your network structure.
For calculate for MSE, your network needs GT. (sharp image)

But i have blurred images only.
Can i try to train your code??

Thanks,
Edward Cho.

how long and how many epochs(~3000?) will the result be reproduced?

The code in the file :
images['gt'] = Variable(inputs['sharp_image'] - 0.5).cuda(GPU)
H = images['lv1'].size(2)
W = images['lv1'].size(3)

Errors outputs
H = images['lv1'].size(2)
KeyError: 'lv1'

hello, thanks for your great work, I want to test my 800x450 image, but some error raises up:
RuntimeError: The size of tensor a (225) must match the size of tensor b (228) at non-singleton dimension 2
Could you tell me how to address this problem? Thanks a lot!

In this paper, F-Norm is used to find the loss between Ground Truth
and output of level-1.

I am curious to know why F-Norm instead of L2-Norm. Is there any
specific reason for using the same? Why not L2-Norm?.

Usually F-Norm is used for matrix and L2-Norm is widely used to find Loss.

It would be great help for me if you could clear my doubts

I have got this error when testing on pretrained model.

File "SDNet4_test.py", line 128, in main
feature[s]['lv2_1'] = encoder[s]['lv2'](images['lv2_1'] + residual[s]['lv3_top']) + feature[s]['lv3_top']
RuntimeError: The size of tensor a (499) must match the size of tensor b (504) at non-singleton dimension 3

你好作者 @HongguangZhang 首先这是我用过的效果目前看最好的库，非常感谢!但是我发现有些图会出现方块一样的暗纹，不知道是为何？

play with a different dataset and it is surprising that DMPHN produces super bad loss with nn.init.kaiming_normal_, xavier_uniform_ is much better, default pytorch initialization is the best

wonder other people also find similar issues

In the test stage,I found the average time to process a 720P image is about 0.91.But the dissertation says 30ms in the 1-2-4-8 model.Wonder why?

dear author:
I have a question that training on gopro, have 2 blur folder, that blur and blur_gamma, I want to know you used is blur or blur_gamma? thankyou

我的数据可能不是按照您这个模型来配置的，请问有完整的说明书参考一下么？

您好，有个问题想问您。我在GoPro数据集的测试集上测试DMPHN时，使用了您官方github给的checkpoint；
在论文中SDNet4.py和DMPHN_1_2_4_8.py应该对应的是Stack(4)-DMPHN和DMPHN(1-2-4-8)，论文中的给的score分别是31.20和30.25
但是我自己跑的score两个都只有27点几，和论文中给score差很多
我想问的是：您给的checkpoint是最优的checkpoint的吗？还是只是没训练完整的checkpoint？
希望能尽快得到您的回复，真心感谢！

DMPHN-v2_1_2_4_8网络在测试时，是否可以只输入blur信息，如果可以怎么样修改测试代码

In your paper the VMPHN has the best performance, but I can not found it in your repository.

Can you provide PSNR and SSIM code ?
I never knew which function to use to calculate SSIM and PSNR. This problem has troubled me for a long time.
THX a lot.

In DMPHN_1_2_4.py line190:
print("epoch:", epoch, "iteration:", iteration+1, "loss:%.4f"%loss.item()), 'time:%.4f'%(stop-start))

you got an extra ')', it should be
print("epoch:", epoch, "iteration:", iteration+1, "loss:%.4f"%loss.item(), 'time:%.4f'%(stop-start))

so does the error in DMPHN_1_2_4_8.py.