yuvalbahat / explorable-super-resolution Goto Github PK

Hello，When I was reproducing your project, the following problem occurred：

Why dataset is str, is there any problem with the code here?
How to solve this? Looking forward for your careful reply, thank you!

Hello!

I'm trying out the CEM in isolation with a regular ESRGAN trained model (using the official pretrain model as benchmark), but I have two questions.

When wrapping the network with CEM, I run out of VRAM, which doesn't happen with the model normally, but I really can't see what operation could be consuming the additional memory. Any suggestion for what could be happening?

Also, I noticed something similar happens in the paper's example, but when compared to the regular ESRGAN model upscale, while the wrapped result does seem sharper, it results in ringing around the edges:

Here left is LR, middle is regular ESRGAN, right is ESRGAN wrapped with CEM.

I imagine this ringing is reduced if the model is now trained using CEM, but same as the previous question, any additional recommendations?

I will test it out with other models besides RRDB as well, its an interesting way to combine internal and external learning, but I would like to find if I'm doing something wrong.

Cheers and thanks for the great work!

Hello again : )

when training and testing with CEM warped SR models, I found that if the input LR image has noise, then the CEM output is also noisy. An example of LR input and HR CEM output:

Such behavior is desirable to keep the fidelity. however, noise exists in many SR settings so it is not desirable to make HR output follow the noise. I am wondering is there something we can do in CEM to avoid the fidelity to noise?

I get this error when trying to launch the GUI on Windows.

Python, PyTorch and CUDA are all installed and up to date.

D:\AI\Explorable-Super-Resolution>python GUI.py -opt ./options/test/GUI_SR.json
D:\Software\py37\lib\site-packages\sklearn\feature_extraction\text.py:17: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3,and in 3.9 it will stop working
from collections import Mapping, defaultdict
Using GPU #0
libpng warning: sBIT: invalid
Adding toolbar Load & Save
Traceback (most recent call last):
File "GUI.py", line 2545, in
window = MainWindow()
File "GUI.py", line 1319, in init
self.setupUi()
File "D:\AI\Explorable-Super-Resolution\MainWindow.py", line 309, in setupUi
layout_cols=4)
File "D:\AI\Explorable-Super-Resolution\MainWindow.py", line 48, in Define_Grid_layout
cur_row,cur_col = self.Greedily_Find_Location(occupancy_map=occupancy_map, button_size=button_sizes[button_num])
File "D:\AI\Explorable-Super-Resolution\MainWindow.py", line 24, in Greedily_Find_Location
occupancy_map = np.pad(np.logical_not(occupancy_map),((0,button_size[0]-1),(0,button_size[1]-1)))
TypeError: pad() missing 1 required positional argument: 'mode'

When I retrained the model, I followed your steps to the last step, but the following error occurred, how to solve this? Look forward to your careful answer, thank you!

thanks sir for interesting paper..I am getting an error related to module jpeg2dct.. can you please let me know. Is it being downloaded from some repository since it don't seem to be pytorch or python function.

How to run the GUI without GPU available

Thanks for sharing the training/testing code first!

I've been tryinng to train the ExplorableSR on some other dataset. I found that the generator_step flag is always False after trainning for a few hours. After some debuging, I think it is due to the discriminator's success rate is lower than the default 0.9 threashold.

I am wondering how many steps are usually needed to get the discriminator reach the 0.9 success rate? It would be helpful to provide a pretrained discriminator to speed up the training process.