Wrong output resolution about deep-exemplar-based-video-colorization

[juliokarl]

Hi.
I made a quick fix for this.
Please note the order of the parameters, first height then width:
--image_size [height,width]

Download the patch below and apply it to test.py this way
patch test.py < patch.txt

patch.txt

[juliokarl]

It has been a long time since i wrote that patch.
As a hint, try a different resolution if you get an error. The width/height has to be dividable by 16, maybe it was a different number.
Hope that helps.

[hrdunn]

@semel1 I had a similar issue, I ended up forking the repo and changing the default value in test.py but then I started getting another error likely related to the size I specified being bigger than the input. I'm not sure why the argument didn't work but tbh I didn't spend much time looking into it. Would love to know the answer to this as well. I though it might have something to do with how the model was trained.

[hrdunn]

@semel1 did you ever make any progress on this? I've been busy with other work but will try to look into it this week.

[orangpelupa]

anyone find out how to change the output resolution? @hrdunn what did you change in test.py?

[UpscaleHistory]

I fixed mine using @juliokarl 's patch. If your os is windows like mine, then the patch command won't work.
You need to manually open the test.py with notepad and paste
if type(opt.image_size) is str: opt.image_size = eval(opt.image_size)
into the file, right after the following lines:

filenames.sort(key=lambda f: int("".join(filter(str.isdigit, f) or -1)))

# NOTE: resize frames to 216*384

Then, you need to replace this line:
parser.add_argument("--image_size", type=int, default=[216 * 2, 384 * 2], help="the image size, eg. [216,384]")
with this line:
parser.add_argument("--image_size", type=str, default=[216 * 2, 384 * 2], help="the image size, eg. [216,384]")

It's best to create a backup of test.py just in case.

[enteranon]

When I change the output resolution with the patch and --image_size I always get this error
Sizes of tensors must match except in dimension 1. Expected size 135 but got size 134 for tensor number 2 in the list.

[enteranon]

Yes, no problem, thanks for the quick reply it works but the vram is immediately busy when I go to 1088*1920. You should choose a lower resolution and then transfer the color information to black and white source frame. This is less computationally intensive and hardly distinguishable from the image quality.

[enteranon]

The following function first scales the low-resolution color image to the size of the original image and projects only the color information onto the Y channel of the black and white image.

colorconvertertosourcebw.txt