raywzy / bringing-old-films-back-to-life Goto Github PK

Great work! Can you provide the test code and pre-training model? @raywzy

this project is extremely impressive . we need colorization code

When I try to test it after I trained it a bit, it gives me that error how much I try.
I extract the frames of the video to 'ata' folder and write that. In the training there is no error but when I try to test it, it always gives me this error.

Hi, it such a amazing work!
I am trying to train the model on another training data, and I am wondering Is there anywhere to get Test, validation set for dataset? I mean the one after added degradations, in Quantitive result in Table 1.
thanks

I am trying to train the model for the colorization task.

How did you train the model for colorization task?

Did you use the pre-trained model of Deep-Exemplar Video Colorization for preparing the dataset of colorization?

Can you provide the pre-trained model of colorization?

Thank you

Currently processed videos get converted to black&white regardless of the color of the sources. Is it possible to retain/enhance original colors in output similar to your Bringing-Old-Photos-Back-to-Life - "Quality Restore"?

Is there any Colab notebook available having the implementation of this Github repo just like the one for the photos.

Hello!
After training, I got several model files in my OUTPUT folder, such as:
net_D_xxx.pth & net_G_xxx.pth & optimizer_xxx.pth

For the next step, how to use these model to colorize/denoise my black&white footage?
(I've converted my footage to be png sequence frame images and put them into a folder already)
Any tips will be helpful!! Thanks!!

In your enclosed demo video I can see noticeable spatio temporal Inconsistency (flickering) in original and processed videos (both restored and colorized)-just take a look at the upper right part (sky) of your video. As I understand restoration process should remove such defects.
Do you plan to resolve this issue?

After processing scratches are not removed

The algorithm works very well, but how to get the same output size as the input image? All output images are (480x368)?
Thanks a lot and congratulations for this great work

Great work. Can you provide the train/test code for the colorization task?

Hey,
did somebody know what the temporal_length, temporal_stride means?
Is there a temp dir for the picts that are already finished, or stay they in ram?

Cheers george

How to generate paired data via the Video Degradation Model? How to use the relabled scratch templates?

Hi, could you explain how you compute the backward and forward hidden states during training?
In the paper, you state that s_(t-1) has aggregated all available information from frame 0 to frame t − 1 but if I understood it correctly this means that for each example you need to precompute the hidden states with all the frames of the video.

Excellent work! When I use your pre-training model to test the test movie you provided, the input_url and gt_url are both in the same folder, but there will be an array out-of-bounds error later. Can you give me some suggestions? thank you！
my order：CUDA_VISIBLE_DEVICES=0 python VP_code/test.py --name RNN_Swin_4 --model_name RNN_Swin_4 --input_video_url ./test_data/data_1/003 --gt_video_url /test_data/data_1/003

I have a nice RX7900XTX and would prefer to not have to go out and buy an nvidia as well. Not enough slots with these new cards.
I'm very interested in restoring some old family film that has been transferred to horrible interlaced 480i. Thanks.

As a result, I get two videos, one original, the other converted. I need to receive only the converted form as output.

The validation set in the profile I'm looking at doesn't seem to be using REDS, but real film video footage, is that correct?

Hey, is there anyone that has the same problem?
The software runs for a couple of folders then stops with that error one a folder.
Cheers george

I have found the checkpoint files in the directory, but I didn`t notice the code or parameter to load the checkpoints. Did I just miss it? Or maybe I asked a stupid question : )

It's been quite a while since the author created this repo. Do you have expected date for code release?

Thanks for open source, Does it needed to use pixel loss training first, then load the pre-training model and add GAN loss to continue training?

Optical Flow works bad in occlusion case after warp

is there anything we can do to reduce this issue?

hello , Dr. Wan,When I run the code from your previous paper《bringing old photos back to life》,it occurs error:

Traceback (most recent call last):
File "train_domain_B.py", line 48, in
model = create_model(opt)
File "D:\pythonProject\6_18\BringOldBack\Global\models\models.py", line 17, in create_model
model.initialize(opt)
File "D:\pythonProject\6_18\BringOldBack\Global\models\pix2pixHD_model.py", line 40, in initialize
opt.n_blocks_local, opt.norm, gpu_ids=self.gpu_ids, opt=opt)
File "D:\pythonProject\6_18\BringOldBack\Global\models\networks.py", line 60, in define_G
netG = GlobalGenerator_v2(input_nc, output_nc, ngf, k_size, n_downsample_global, n_blocks_global, norm_layer, opt=opt)
NameError: name 'GlobalGenerator_v2' is not defined

how did you solve it?could you please share with me ? thank you.

Hi, congratulations for this work, I could not execute the scripts because I get this error:

FileNotFoundError: [Errno 2] No such file or directory: './configs/RNN_Swin_4.yaml'

is there any video tutorial? I think we need a tutorial.

Upon launching :
set CUDA_VISIBLE_DEVICES=0 & python VP_code/main_gan.py --name RNN_Swin_4 --model_name RNN_Swin_4 --epoch 20 --nodes 1 --gpus 1 --discriminator_name discriminator_v2 --which_gan hinge

it tries to establish an unsolicited connection:

[W ..\torch\csrc\distributed\c10d\socket.cpp:601] [c10d] The client socket has failed to connect to [link.redgiant.com]:28172 (system error: 10049 - The requested address is not valid in its context.).

P.S Due to incapability Windows 10 with "nccl" I changed line 41in "main_gan.py" from
backend = 'nccl' to backend = 'gloo'