DeblurGAN icon indicating copy to clipboard operation
DeblurGAN copied to clipboard

Are the generator weights are only kept in the "test" runtime code in this repo or We have to take care of it?

Open abhishekvahadane opened this issue 6 years ago • 5 comments

  1. "Download weights from Dropbox . Note that during the inference you need to keep only Generator weights". Are the generator weights are only kept in the "test" runtime code in this repo or We have to take care of it?

  2. Also once we run test.py, it generates the results folder with "fake" and "real" image. I undestand the fake as a deblurred image. Is that right?

  3. The result images generated are below. The fake_b is still blur. Any problem with the dropbox weights? Anyone got the same results?

Real: yolo_b_real_a Generated Fake: yolo_b_fake_b

@KupynOrest

abhishekvahadane avatar Apr 04 '18 10:04 abhishekvahadane

I'm having the same problem. Have you figured out what's the issue?

rachel-rj avatar May 07 '18 07:05 rachel-rj

You need to put the Generator weights into the appropriate place and load the model in test time. After that, the 'fake' output should be deblurred version of the image. If you would experience any problems with that process, feel free to send me the message

KupynOrest avatar May 07 '18 12:05 KupynOrest

Hi @KupynOrest, thanks for the quick reply. After changing the option --resize_or_crop to no_change in order to keep the output file as the same size as my input files, I find your code works for images with pixel sizes less than 256px. When I tried with your example image without any cropping, the output is still blurry like what @abhishekvahadane showed. I've included the results below as well. First image is the input, second image is the output of the GAN.

yolo_b_real_a

yolo_b_fake_b

rachel-rj avatar May 08 '18 07:05 rachel-rj

Hi @KupynOrest, thanks for the quick reply. After changing the option --resize_or_crop to no_change in order to keep the output file as the same size as my input files, I find your code works for images with pixel sizes less than 256px. When I tried with your example image without any cropping, the output is still blurry like what @abhishekvahadane showed. I've included the results below as well. First image is the input, second image is the output of the GAN.

yolo_b_real_a

yolo_b_fake_b

I have the same problem with you,the image of result is still blur,just like the generator doesn't work

Maxairw avatar Nov 27 '18 12:11 Maxairw

@KupynOrest thank you for the project first,then i need to change a blur image to be clean,but when i run the test.py,the result has two images .one is real,another is fake,but both is blurred,and the size is changed to 256*256,i do not know the reason,please give us a guide,big thanks.

boyang24 avatar Feb 26 '19 06:02 boyang24