aiff22 / pynet Goto Github PK

Google Drive returns:

Sorry, you can't view or download this file at this time.

Too many users have viewed or downloaded this file recently. Please try accessing the file again later. If the file that you are trying to access is particularly large or is shared with many people, it may take up to 24 hours to be able to view or download the file. If you still can't access a file after 24 hours, contact your domain administrator.

I do not know how to get this.

Thx.

which one is right？

Hi, because the datasets too big, I can't download it fully from goole driver, can you split into small pathes?

Hi, Where can I find the EBB complete dataset?

The competition site, to which I'm directed to get the test dataset at-leats to is not updated since 2020. So it does not provide the complete dataset.

This is for a university project. Would be appreciated very much if this could be addressed ASAP.

I want to know how to get the Bayer Photo by myself with HUAWEI or iPhone or Other Android phone.

I tried on a 448*448*3 PNG image but this error occurred :

Traceback (most recent call last):
  File "test_model.py", line 67, in <module>
    I = np.reshape(I, [1, I.shape[0], I.shape[1], 4])
  File "<__array_function__ internals>", line 6, in reshape
  File "C:\Users\127051\AppData\Local\Programs\Python\Python37\lib\site-packages\numpy\core\fromnumeric.py", line 301, in reshape
    return _wrapfunc(a, 'reshape', newshape, order=order)
  File "C:\Users\127051\AppData\Local\Programs\Python\Python37\lib\site-packages\numpy\core\fromnumeric.py", line 61, in _wrapfunc
    return bound(*args, **kwds)
ValueError: cannot reshape array of size 602112 into shape (1,224,224,4)

Such as DNG photos taken by Quad-Bayer or RYYB filter? Thanks!

Hi andrey

Thank you for share PyNET source code and pretrained models !

I sent you an email(to [email protected]), but unfortunately, for now you have not reply me.

I have cloned PyNET code and downloaded data set and pretrained models.
https://github.com/aiff22/PyNET

I downloaded a dng pictrure(get form LEICA M9) form : https://www.kenrockwell.com/leica/m9/sample-photos-3.htm.
I test the pictrure on your pretrained PyNET model, unfortunately I got a very bad image quality.

I provide my dng pictrue(L1004235.dng), input png picture(get by dng_to_png.py(L1004235-input.png)) and output png(L1004235_level_0_iteration_None.png) as attachments in the mail I sent to you.

Hope you can help me to fix this issue.

Hello, Why is the calculation of loss not consistent with the paper, and the weight of each part of loss is not the same? The tensorflow version.
Look forward to your reply. Thanks.

Dear author
Hello
I hope you will grant my request to download your pre-trained PyNET model

I encountered an issue when training at the last level. When I execute the command

python train_model.py level=0 batch_size=10 num_train_iters=100000

I got the following error:
Loading training data...
Killed

Any ideas?

As far as I am concerned， the huawei P20's raw image is 10bits dng files. but the data downloaded form the link http://people.ee.ethz.ch/~ihnatova/pynet.html#dataset are mixed with 8bits and 10bits png files. And the normalization code for training and testting in load_dataset.py at 21th line is shown below where the divisor is 4*255:

RAW_norm = RAW_combined.astype(np.float32) / (4 * 255)

Since there are part of data are 8bits depth, the divisor is too big for them.

And I tryed to used the pretrained model to test my own data captuered by huawei P20, the result is slightly overlighted.

So Is that a bug？ or something far from my understand?

I try to fix this by replace the code in load_dataset.py at 21th line and re-train the model:

RAW_norm = RAW_combined.astype(np.float32) / (4 * 255)

by

    # 10bits

   RAW_norm = RAW_combined.astype(np.float32) / (1023)

elif raw.dtype == np.uint8:

    # 8bits

    RAW_norm = RAW_combined.astype(np.float32) / (255)

hello, does this can be trainned on mutil gpus?
when i train these code with two gpus, while just exe on one gpu.
not familiar with tf, could you please give some guidence how to trainning on mutil gpus with your code? thanks.

Hi:
Thanks a lot for sharing your source code.
In your paper, the loss for level 1 is loss_1 = L_vgg + 0.75L_ssim + 0.05L_mse, and each loss is normalized to 1. While in your code, I think the condition of LEVEL ==0 is same as loss_1 in your paper. The loss is loss_generator = loss_mse*20 + loss_content + (1 - loss_ssim)*20.
Q1, why are they different? How to normalize L_ms to 1?
Q2, why is these weights for loss_mse and (1 - loss_ssim) the same. I think loss_mse is a big value(>50), while 1-ssim is smaller than 0.1.
Q3, could you tell me the value range of these 3 losses? Maybe this can help me to understand the weights for loss. In my case, the loss_mse is around 100, ssim is around 0.98, and loss_content is around 7.
Thanks in advance.