We use your code to read the EXR file in the dataset and find that many of them have NAN values. Could you tell me how to right read EXR file or deal with these NAN values?

Dear Shangchen Zhou,
for my research scope, I needed to infer the results produced by the single DeblurNet. I followed the procedure described (here) and was able to switch to do the Single Deblurnet configuration. I shall be grateful if you kindly help me to get the checkpoint of the single deblurnet. Specifically, I am looking for the pre-trained model which is referred as "Ours-Single" in table 1 of the paper.

Thanking you,
Oaishi.

I noticed that the disparity map in the dataset is saved in EXR format. How can I export it to PNG format

Dear Zhou:
I run the training code and find that the epoch num is 400 in config.py, while 2000k iterations is proposed for the deblurNet. Are 400 epochs sufficient for training deblurNet?
Another issue is how to train with FlyingThings3D is not presented clearly in code, thank you for more clear guidance~

good job!.do you plan to release the code?and when?

Hi DAVANet Team,

Can you please share the pre-trained model so we can test on it

in build.py line124

deblurnet_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(dispnet_solver,milestones=cfg.TRAIN.DEBLURNET_LR_MILESTONES,gamma=cfg.TRAIN.LEARNING_RATE_DECAY)

why 'dispnet_solver' appears here....
thanks for much

Hello, thank you so much for the great repository. I wanted to know about the procedure of your dataset preparation. Did you use the default calibration settings of ZED camera? Or redefined the calibration yourself?

Hi I am following up the question I asked earlier in here:

#6

But I am specifying 'DeblurNet' instead of 'StereoDeblurNet' in the configuration file, and still the network complains about the tensor size mismatch. As from your Table1, it feels you have two different models for deblurring; one using the fuse information from fusionNet and DispBiNet while other not using this information. Is the trained model doing mono deblurring available?

Also, can you please not close an issue as someone could have follow up questions.

Hello, your work is really amazing!

I have some questions.On Gopro datasets, I used a single deblurnet(yours)to train the parameters. The experimental setup is that the train_batch_size is 2, the test_bs is 1, the initial lr is 1e-4, and lr is decayed by 0.1 in [80, 160, 240], and the other parameters are consistent with your code.

I have a few problems :
(1) Such a small batch, Memory-Usage and GPU-Util are small, is this a waste of resources? I ran 200 epochs in a day and a half, and the PSNR only reached 28.8 not 30.55 reported in your paper, but there is no sign of rising.
(2) If I use a large batch_size, the PSNR performance is worse.

What do I need to pay attention to in terms of training speed and code tricks to reproduce the results of around 30.5 in your paper? If you can give some guidance, I really appreciate you!

Thanks for your nice work again!

when I set the phase to test,I get the following result,I could not find the deblur result,and the deblured picture's color strange,what should I do?

left picture is original picture,the middle picture is original sharp picture,and the right picture is deblured picture use your given model{ckpt-best.pth.zip}

Excuse me, what is the PSNR value of the first 10 epochs or so on GroPro with single DeblurNet?If give me a reference, I really appreciate it.

我这里用的torch版本高于1.1.0，所以报出了warning。我想根据warning去调整lr_scheduler.step()和optimizer.step()的位置，但是我一直找不到optimizer.step()的位置

hi
thanks a lot for the great project, looks amazing.

Can you help on how to run the inference on images in the wild with the pre-trained model?
Do I need to modify the test code to do so or is there some code to do that already in the repo?
What data do I need to give as an input to the model, just the left/right image pairs or do I need to pass a depth map as well?

Hello,

Thank you for open-sourcing the codebase. I was trying to test your algorithm with the option deblurnet and I face the following error :

Traceback (most recent call last):
  File "runner.py", line 56, in <module>
    main()
  File "runner.py", line 53, in main
    bulid_net(cfg)
  File "/home/DAVANet/core/build.py", line 106, in bulid_net
    init_epoch = checkpoint['epoch_idx']+1
KeyError: 'epoch_idx'

Can you help fix this? Much appreciated !

1)how could I train gopro dataset?
2)now,I am working with car card deblur,does your work is fit in my project?
3)how about the GPU memory usage?and the test speed in gopro dataset?
Thanks

您好，在原论文中，DeblurNet第13-14的子块是AtrousResBlock，但是在开源的代码中使用的ResBlock，是不是开源时代码不小心错误？
如果黄色子块确实是AtrousResBlock而不是ResBlock，可以给出AtrousResBlock的结果吗？因为您在论文中没有描述AtrousResBlock的结构
如果能够解答，真心感谢！

Hello, I am trying to run the inference using the deblurnet module of DAVANet. Have changed the flags appropriately in the config file. But, when it goes to build_net(cfg) -- it gives me the following error:

"size mismatch for module.upconv3_i.0.weight: copying a param with shape torch.Size([128, 288, 3, 3]) from checkpoint, the shape in current model is torch.Size([128, 128, 3, 3])."

It is somehow complaining about the mismatch in the tensor sizes but I am not sure why this could be. Please advise!