yzhouas / pd-denoising-pytorch Goto Github PK

Can you send me the dataset Set20 ？ my email: [email protected] .Thank you very much!

I realize that:

"This version of testing script may cause exceeding memory issues in GPU while testing on large images. A testing_on_full image version will be released soon."

But I encountered this issue on an 8GB RTX 2070 on ubuntu 18.04 while trying run_test_on_real_patches.sh

I can confirm that it runs just fine when I resize the provided sample images to 300x300px.

Does this seem right? Will a new version of the script solve this problem? Any idea how soon that might be released?

Thanks!

I want to verify the results in Table 2 of paper, so i want to know the parameters in test.py.
my settings for the model, logs_color_MC_AWGN_RVIN, are
python test.py --scale 1 --ps 2 --ps_scale 2 --real_n 0 --spat_n 1 --k 0 --mode MC --color 1 --output_map 0 --zeroout 0 --keep_ind 0 --num_of_layers 20 --delog logs/logs_color_MC_AWGN_RVIN --cond 1 --refine 0 --refine_opt 1 --test_data CBSD68 --out_dir results/CBSD68
my result is 29.1037 when s=20 and c=10

Hi, I have read your paper. It' cool to use this pd denoiser. But I cannot understand the edffect of the 3,4 step, you just use this and say these steps can get the texture regions. Can you explain the effect of 3,4 step. Thanks

Hi! I uncommented the "eval" part in the training loop, and want to use it for validation during training (with color images).
However, I get one of two errors:

Because of the line:
noise_map_val = np.zeros((1, 2*c, img_w, img_h))

In the code

The marked line causes a dimension error, as it tries to input a 3 channel image into a 6 channel tensor (as defined above with 2*c).

Otherwise, if I change the above line to
noise_map_val = np.zeros((1, 1*c, img_w, img_h))

The line
out_val_nb = torch.clamp( imgn_val - model(imgn_val, NM_tensor_val), 0., 1.)
Fails, because NM_tensor val has 3 channels and not 6.

What is the right way to input 6 channels into the generated noise map from 3 channels?

Thanks,
Tom

Hi,

Thanks for your great work on the real-world denoising!
It gives me lots of intuitions.

It's a simple bug while using non-zero k and pre-set stride mode(ps=2).
The oversmoothed image does not be reverse pixel-shuffled.

I think the problem on the N173 in test.py
: max_out_numpy = visual_va2np(max_Out, opt.color, opt.ps, pss, 1, opt.rescale, w, h, c)

I changed it to:
max_out_numpy = visual_va2np(max_Out, opt.color, opt.ps==1 or opt.ps==2, pss, 1, opt.rescale, w, h, c)

Is this right modification for your intend?

Thank you

I ran test.py without reporting an error. I set it exactly as the README says. It is normal to read the model and pictures, but only the result saving path is generated, but there is no picture generated in the path. Can you tell me what is the reason?

hi, i find this problem, i don't understand how give it solution.
@yzhouas

Hi,
In Figure 3, the basic model is very similar to the CBDNet architecture, but in Table 3 there is a big gap of the PSNR values between CBDNet and "our basic model". I want to know what difference causes this phenomenon. Can you help me find some clues?

Thanks!

I just simiply run bash run_test_on_real_patches.sh .
hard_human_512_pss2_k0.0.png

crop8_pss2_k0.0.png

can you please let me know how to test the model on cpu because it asks for a gpu everytime i run the script