hhb072 / waveletsrnet Goto Github PK

What are the typical texture loss values while training... I have done my own implementation of the above network.. the texture loss has a torch.mean its value to 0.0005 range.. where as other losss sr loss, lr loss and full image loss have a torch.sum() component.. and are just divided by the batch size... they are in 10^3 range the total loss is dominated by sr loss,lr loss and full image loss. I dont understand how the texture loss is going to contribute towards the learning.

I trained the model by myself,but the results were terrible,there are always black things around eyes.

could anyone else help me ,i fail to run this code . it shows in this part of code:
loss_lr = loss_MSE(wavelets_predict[:,0:3,:,:], wavelets_lr, opt.mse_avg)
loss_sr = loss_MSE(wavelets_predict[:, 3:, :, :], wavelets_sr, opt.mse_avg)
loss_textures = loss_Textures(wavelets_predict[:, 3:, :, :], wavelets_sr)
loss_img = loss_MSE(img_predict, target, opt.mse_avg)
the "wavelets_predict[:,0:3,:,:]" tensor.size ([16, 48, 128, 128]), but "wavelets_lr" is([16, 3, 32, 32]), i wanna know did anyone meet the same problem, how did you solve it?

thx for sharing
in ur paper, u mentioned that u used sgd
but ur code used adam instead
i changed sgd to adam, but the model couldn't get converged, why did it happen?

Hello, I got this error in the python3 version of the code. Can anyone help me?

I am trying to running the code. It shows an error: No such file or directory: 'path/celeba/train.list'
which mean I have to get the CelebA dataset first. But how should I get the dataset and how should I matching the directories?

Hello, could you please provide your training model file or checkpoint ? I want to test it on my data set

Hello, authors. Thanks for your contribution and good approach.

I have tried to test your code by making test only code and today fully tested you network.
I trained your network (for 950 epoch) and then tested several images by importing the trained model.
Then I found very good result for some images but sometimes I saw very strange noise.
One of the most serious artifact on reconstructed images was like a small snake (let me call it small snake noise, SSN).

When the network has very clean image, the performance was excellent!!
However, when I input a noisy image such as jpeg noise, in the reconstructed images, I could see that noises.
It look like trying to connect every tiny details (some times noises), so very small noises were connected one another and boosted up. In my opinion, they originated from Wavelet method and hence I thought it can be improved by adjusting Wavelet method or something. Could you please tell me how can I or please guide me if I have a wrong point for the problem.

Thank you so much.

After reading your Wavelet-SRNet paper,there are three questions in the paper requires consulting you:

1. How is the channel number 1024 of the output feature map after the embedded network?
2. Why is the number of channels in the wavelet prediction network 32 and 64?
3. Can this reconstruction method improve low-resolution face recognition rates?

I spent one whole day training it on school's server. But it just finished one epoch.Is it right?

I have trained the net with the Celeba datasets,but the result is not so satisfactory. Could you please upload the pretrained file/weight file ? Appreciate your help

Issued in interest by a fellow developer.

Hi,

It seems the uploaded model doesn't work. is it possible to update it? Thanks!

Hi, hhb072!

I'm trying to implement your project, but some problems occurred. That is I don't know how the data is arranged in your demo. So would you submit an example for that?

Thank you!

Can someone provide the wavelet weights for x8 upscale? It's not included in the provided pkl file (the key is only up to rec16).

Thank you for sharing your code.

But, could you tell me in detail, how can I test the SR network without training?

I'm so sorry but I'm not familiar with these kind of code such as tensorflow and python.

Please describe the way. Thank you in advance.