yzxing87 / pytorch-deep-video-prior Goto Github PK

Hi,
I see that the loss used for IRT in the code is different than the one in the paper.
Can u pls explain why is there a difference pls?

Also, i see u divide the pixelwise difference of O_main and the net_gt by the image. What's the idea behind this?
Thank you!

Upon running "main_IRT.py", got an error:

Exception has occurred: ValueError
attempt to get argmax of an empty sequence
  File "D:\colorization\deep-video-prior-master-pytorch\main_IRT.py", line 49, in <module>
    for x in subprocess.Popen("nvidia-smi -q -d Memory | grep -A4 GPU | grep Free", shell=True, stdout=subprocess.PIPE).stdout.readlines()]))

Happened in both Windows 10 and WSL2
Nvidia RTX 3090

Hey,
this is a great project!!! What is the license of the code?

Is it possible to generate temporally consistent video from depth map sequence obtained by depth estimation (MIDAS) similar to "Blind Video Temporal Consistency project" https://github.com/lulu1315/BlindConsistency or https://github.com/nbonneel/blindconsistency?

When attempting to run main_IRT.py with 6000 frames causes RuntimeError: CUDA out of memory. This could be fixed by reducing batch size or use memory when needed. Also, I`m using google Colab and yes the ram that google collab had offered did not reach the full amount of ram instead, it stopped at 2/3 of it and caused this error

Traceback (most recent call last):
File "main_IRT.py", line 139, in
net_in = torch.from_numpy(net_in).permute(0,3,1,2).float().to(device)
RuntimeError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 14.73 GiB total capacity; 12.13 GiB already allocated; 5.88 MiB free; 13.79 GiB reserved in total by PyTorch)

Hi，首先非常感谢您愿意给出pytroch版本的DVP，对于学习DVP真的帮大忙了！
在复现您论文结果的过程中，发现DVP输出结果尺寸与原图像序列不一致。
例如，在demo中的上色序列为854 * 480，而test.sh输出结果为 832 * 480。在demo其他任务中也是这样。

请问这种情况正常吗？
因为我希望计算结果的PSNR值，但是被这个问题所困扰着。希望您能为我解答这个问题，非常感谢！

I noticed that a similar issue has been closed.
However, I have the same problem.
So, "You can also load the frame just before each iteration. I think torch.utils.data.DataLoader can be a better option for this issue. We will update the code using pytorch dataloader."
Any code update?
Could you please provide me with more detail on how to " load the frame just before each iteration"

Hello authors, thank you for your great work!

I am writing a blog post about your work and I was interested in reproducing the results completely (on DAVIS/the Bonneel dataset). Would it be possible for you to provide evaluation code or instructions as to evaluating E_warp and F_data for your method on these datasets?

All the best,

Is it possible to run this code using CPU only?
Tried "conda install pytorch torchvision torchaudio cpuonly -c pytorch" - no sucssess.
Running Windows 10.

@yzxing87 @ChenyangLEI Hey. Could you please share the script for creating the dataloader used for training ?

When attempting to run main_IRT.py with 6000 frames causes RuntimeError: CUDA out of memory. This could be fixed by reducing batch size or use memory when needed. Also, I`m using google Colab and yes the ram that google collab had offered did not reach the full amount of ram instead, it stopped at 2/3 of it and caused this error

Traceback (most recent call last):
File "main_IRT.py", line 139, in
net_in = torch.from_numpy(net_in).permute(0,3,1,2).float().to(device)
RuntimeError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 14.73 GiB total capacity; 12.13 GiB already allocated; 5.88 MiB free; 13.79 GiB reserved in total by PyTorch)

@ChenyangLEI @yzxing87 Is it possible to get temporally consistent image inpainting using DVP since the original input frames are actually occluded?