haoyuc / maskeddenoising Goto Github PK

Hi, authors, great work!

Does the results in paper is produced by this configuration? Should I train it with 4 GPUs to achieve same results? 1 GPU or 4 GPUs, which one is official setting?

Currently, I train models with 1 GPU and 200K as the paper stated, however, I found disable input masking achieve better results in validation set you provided. Specifically, for SwinIR, the validation PSNR degrade slowly, for SwinIR with masking training, the validation PSNR grows slowly. See the figure below:

Will the code be released soon? Thank you so much for your very significant contributions!

How to get the pretrained models?

Confusions about Masked Operations:

(1) Are masked operations necessary for both pixel and attention maps during both training and testing stages?

(2) According to Table 1 of the original paper, masked operations appear to be more important. What distinguishes the masked operations on attention maps from Attention Dropout operations? Will Attention Dropout achieve comparable performance?

Hi,Bro,thanks you for your code.
can you tell me the command for image denoising?
for exemple, i have a noise image, i have no clean image, now, i wanna to denoising.
your command:
python main_test_swinir.py
--model_path model_zoo/input_mask_80_90.pth
--name input_mask_80_90/McM_poisson_20
--opt model_zoo/input_mask_80_90.json
--folder_gt testset/McM/HR
--folder_lq testset/McM/McM_poisson_20

but, i have only one noise image, if i use:
python main_test_swinir.py
--model_path model_zoo/input_mask_80_90.pth
--name input_mask_80_90/McM_poisson_20
--opt model_zoo/input_mask_80_90.json
--folder_gt testset/mynoiseImg
will be output some fault

It is mentioned in the paper that "At this time, due to the inconsistency between training and testing, the network will tend to increase the brightness of the output image." What can cause the inconsistency between training and testing?

What is the purpose of argument 'value' in input_mask function?

def input_mask(image, prob_=0.75, value=0.1)

It seems the inconsistency between training and testing can be reduced if the 'value' is set as 0.0.

Thank you for the inspiring work. I would like to know how you computed the CKA similarity between dense features of two different noisy images obtained from the same latent image. Do we need multiple samples to compute the CKA similarity?

函数input_mask_with_noise好像是生成一个经过裁剪附加上mask的图像块，请问这是原论文中所指的mask token吗？那么利用mask token随机替换原图像中像素的代码在哪呢？十分感谢！

Hello! I'm really excited about your work for generalizable image denoising. I can't wait to see and run the code. Will the code be released soon? Thank you so much for your very significant contributions! :)

谢谢！

I have seen one paper using the crop and paste operation(crop some part from one image and paste it on the origin image) to enhance the performance. Im not sure if it works in this situation. Just a possible solution~hhhh

Hello, what you have done is incredible, I cant wait to learn and run your code, when will you release the code? Thank you for everything you have done! :)

1. baseline model是否指的就是图4的网络结构去掉input mask和attention mask？如果是，我是不是可以认为baseline model是在swinir的基础上修改后的模型？如果是，你们为什么不直接在swinir上使用input mask和attention mask？
2. 附加材料里说dropout模型是before the output convolutional layer of the baseline model，是在每个block的输出层前面 ，还是在整个网络的输出层前面？

Thanks for your code sharing, Could you tell me about the input mask and attention mask code?

nput_mask_80_90.json 参数里dataroot_H 是无噪声数据集路径吗， "dataroot_L"是退化数据集路径吗

您好！请问您提供的权重（在model_zoo中）是只能处理gaussain noise 吗？如果我想处理其它噪声，我是否需要重新训练？