我刚开始只使用了SIDD数据集进行训练，大概有16000张图片，psnr最终停留在32.x就不再上升了。后来又去给图片添加噪点，就是使用普通的高斯噪声，加了大概3万多继续训练，现在psnr只能达到34.97很难上去了。我想到可能是没有使用DND dataset，只是自己添加了随机的噪声才会导致这样，想请问我理解的正确吗？

原谅我不能使用英文和你交流。

Hi,

I got

NameError: name '_file_' is not defined when running through

opt['path']['root'] = osp.abspath(osp.join(_file_, osp.pardir, osp.pardir, osp.pardir))

at codes/options/options.py.

Any suggestions?

yours, Luming

In the yaml file, 2 folders have to be set:

dataroot_GT: '' # path to training Clean images
dataroot_Noisy: '' # path to training Noisy images

Show I put original SIDD GT and noisy png files into these folders or show I run your crop_sidd.py to create a large number of cropped images? In addition, what should be the file names in those folders?

Hi, I am interested in your paper's title and I wonder when you would release the full paper. Besides, the image link of network architecture in your readme has failed, so I cannot get a brief knowledge about your work.

Thanks in advance.

Hi, thanks for this great works! I try to train invdn in my own datasets. I get a strange problem. The l_back_rec is too big, eg 1e+3. I don't know whether it is normal. I would apprecaite it if someone could help me.

感谢你的回复！不过我有遇到了新的问题，我不知道如何使用自己的测试集生成"BenchmarkNoisyBlocksSrgb.mat”。我在bing搜索了一下，看到了找到了这个链接:"https://github.com/TLampr/VDNet-TF-2.0/tree/master/datasets"，请问是这样生成的吗？如果不是的话，能不能把有关的资料发给我去了解一下。谢谢！

I tried to perform training on DIV2K dataset with the help of LQGTSN_dataset.py. However, the performance seems poor(under 30dB PSNR, quite lower than proposed). The loss curves also differ a lot from the training on SIDD dataset. Maybe I missed some important configurations when training on the DIV2K. It will be appreciated if you provide corresponding pre-trained models or more detailed instructions for reproduction. It seems empty in crop_SIDD.py. Besides, crop_DIV2K.py doesn't match the DIV2K dataset.

The original SIDD dataset is not a folder structure like the author's /SIDD_Medium_Srgb/GT and SIDD_Medium_Srgb/Noisy, but a SIDD_Medium_Srgb\Data\0001_001_S6_00100_00060_3200_L*GT*.png or NOISY.png like this. Please tell me how the author runs the code.

Hi! This is a really interesting work in image denoising, and I try to reproduce the results using your code. However, I have some questions about the details of your code:

1. The train_InvDN.yaml file is different from the description in the paper, e.g., batch size of 28 in yaml vs 14 in the paper, learning rate steps 100k in yaml vs 50k in paper. Which configuration should I use to reproduce your result in the paper? or Could you please provide the original yaml file of your paper?
2. About the loss function: the backward loss still contains the gradient loss and the SSIM loss. Should I comment them or not?
3. Just to make sure how to prepare the SSID training data. Is it right to crop each training image into 100 patches with the patch size of 512 using Crop_SIDD.py?

Thanks for your patience to read my questions, and I'm looking forward to receiving your reply!

Hello, this is a very worthwhile paper, but I have a question for you now. Can you provide the code for the "PCA Visualization Results" part of your paper?Thank you so much！

NICE work~ I tried to use thop on your code to calculate FLOPs, but there will always be errors. Could you please provide the code you report to calculate flops

Hi, thanks for your great work and sharing. I tested the pre-trained model with my test image, and it generated some artifaces. Have you ever encountered this? Any suggestions for improving this?

Hello in line 12 of Denoised_LR.py file there is "from codes.models.modules.VDN import VDN as DN" but in there is any VDN module or function so this generate me an error while trying to denoise an image. Hope I can have an answer, thank you.