sreyas-mohan / udvd Goto Github PK

Hi,

Is there any chance you could share all required packages to run everything contained in the repo (code + jupyter notebooks), either through a simple dotted list, or with a .yml conda environment file (e.g. as done here: https://github.com/m-tassano/fastdvdnet/blob/master/requirements.yml)?

Thank you and congratulations for your very interesting publication, Antoine

I encountered a bug when running single_train.py using the setup in readme
'--model blind-video-net-4
--data-path dataset/Set8
--dataset SingleVideo
--dataset-aux GoPro
--video rafting
--aug 2
--sample
--heldout
--batch-size 8
--lr 1e-4
--num-epochs 32
--step-checkpoints'
The error content is “
File "F:\xxx\udvd-main\data.py", line 534, in getitem
noisy_img = np.load(self.noisy_files[index+i-off])
IndexError: list index out of range”

I think it's because the code
self.len = self.bound = len(self.files) if self.heldout: self.len -= 5
Maybe self.bound should be re-assigned after changing self.len. I changed the code as '
self.len = self.bound = len(self.files)
if self.heldout:
self.len -= 5
self.bound = self.len'
Luckily, the error disappeared. I want to know if my modification is reasonable.

Thank you for your impressive work.

Recently, I have learned your training code and noticed that the network outputs an est_sigma. I'm not sure what this is for? Or if this is used to predict the noise variance.

Further I would like to ask, if I want to use your blind-spot architecture, do I just need to replace my convolutional layers?

Looking forward your reply!!!

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?

Hi!
I am trying to run model training, always get this error:

File "fluoro_train.py", line 202, in
main(args)
File "fluoro_train.py", line 16, in main
utils.setup_experiment(args)
File "C:\Users\annel38\Desktop\udvd-main\utils\train_utils.py", line 53, in setup_experiment
os.makedirs(args.experiment_dir, exist_ok=True)
File "C:\ProgramData\Anaconda3\lib\os.py", line 221, in makedirs
mkdir(name, mode)
OSError: [WinError 123] The filename, directory name, or volume label syntax is incorrect: 'experiments\blind-video-net-4\blind-video-net-4-BF-5-50-Dec-05-20:58:57'

Hello,
sorry but I'm not sure I understand the arguments when creating a UDVD model. Indeed, in your code it is written that BlindVideoNet takes: input channels_per_frame and out_channels.

When I create the model I do UDVD=BlindVideoNet(3,9), because I have three images and the images have 3 channels per frame and in total it's 9 channels. The tensor shape I give to UDVD is [8, 9, 512, 512] for [Batch_size, sequence_total_channel, H, W].

However, your code returns an error telling me RuntimeError: Given groups=1, weight of size [48, 9, 3, 3], expected input[8, 6, 515, 514] to have 9 channels, but got 6 channels instead.

So I'm wondering what is the shape of the tensor that your network takes as input and what is the shape of the output. Thank you!

Kind regards,
Sébastien de Blois

Hi, thank you for your amazing work!

I do not catch the implementation of asymmetric convolutional filters that are vertically causal. Maybe I miss some information...

Could you please give me some hints about implementation the asymmetric convolutional filters that are vertically causal?

Thanks a lot

Hi thanks again for your great work!

I have a question regarding the speed during the training phase I found it a lot slower (5x) than an equally complex FastDVDNet model, UDVD has almost the same architecture as the FastDVDNet, so why is that? I know that the training scheme is completely different cause UDVD is self-supervised and FastDVDNet is supervised, but still it is a huge speed difference between both methods.

Thanks a lot for your help.