wenchao-du / lir-for-unsupervised-ir Goto Github PK

As your attempt, you want to divide the pascalvoc dataset into 4, and this .py file is just for pascalvoc dataset.
But in your code :
def splitlist(folder, trainfolder1, trainfolder2, testfolder):
if not os.path.exists(folder):
raise Exception('input folder dose not exist ! please check it !!')
if not os.path.exists(trainfolder1):
os.makedirs(trainfolder1)
if not os.path.exists(trainfolder2):
os.makedirs(trainfolder2)
if not os.path.exists(testfolder):
os.makedirs(testfolder)

filelist = os.listdir(folder)
sublist = random.sample(filelist, **110000**)
sublist_a = random.sample(sublist, **55000**)
   .....

The number of pictures in the folder is 110000 and 55000 respectively, but there are only 12031 images in the pascalvoc data set. How did you do it? Or is it that you made a mistake?

The work is interesting! However, in your paper, you only add a KL divergence loss to regularize the distribution of the noise code, while you add KL loss to all latent features in your open source code. Why is there such a difference? Is KL loss important to the final result？

Hello, I want to ask what is '/mnt/B290B95290B91E33/Dual_UNIT/outputs/unit_noise2clear-bn-Deblur/checkpoints/gen_00300000.pt' in test.py

I just test some effects of model, so can you offer your pretrained model?

in dis_update(self, x_a, x_b, hyperparameters):
the content adversial codes are ,
out_a = self.dis_content(h_a)
out_b = self.dis_content(h_b)
why not use h_a.detach() and h_b.detach() to avoid the update in the parameters of G ?
Thank you!

我的数据集是640*512的红外噪声图像（Celeba_B）与干净图像（Celeba_A），但是一直在 File "G:\WIN\LIR-for-Unsupervised-IR-master\networks.py", line 344, in forward
return self.model(x)
与 File "G:\WIN\LIR-for-Unsupervised-IR-master\networks.py", line 467, in forward
x = self.conv(self.pad(x))的 def forward(self, x):
x = self.conv(self.pad(x))
if self.norm_type != 'wn' and self.norm != None:
x = self.norm(x)
if self.activation:
x = self.activation(x)
return x
报错 return F.pad(input, self.padding, 'reflect')
TypeError: pad(): argument 'input' (position 1) must be Tensor, not NoneType
实在毫无头绪，恳求学长帮助

您好，能给一个更详细的训练过程吗？就是每次该运行哪些.py文件

Hello
How are you?
I am going to apply this method to a general image restoration such as deblur & denoising.
I found that the training step requires testA and testB data too.
Are these testA and testB data unpaired data?

It is expected that other types of noise are possible to be learned by the network, for example, JPEG artifacts?

Does the clean data set come from the VOC data set? So do you need to search for noise data by yourself? Where is the noise data set used by the author? Is the clean data set put into Celeba_A, and then the noise data set into Celeba_B?

Hi. The paper says there are two encoders (generators) for the clean image and noisy image splitly. However, I can only find one generator 'gen_a' in 'train.py'. I think that it is the generator for noisy image. Where is the defination of the other generator (for clean image)? Thansk in advance.