Missing modules data.py, denoisenet_full.py, and config_refine.py

[iansampson]

Thanks for posting this code! Really fascinating paper.

I tried running denoise_main.py in Colab and ran into some errors. A couple Python files seem to be missing. Here’s all my Colab code:

import os
!git clone https://github.com/yuguochencuc/SF-Net.git
os.chdir("SF-Net")
!python denoise_main.py

And the error that comes back:

Traceback (most recent call last):
  File "denoise_main.py", line 3, in <module>
    from data import TrainDataset, CvDataset, TrainDataLoader, CvDataLoader
ModuleNotFoundError: No module named 'data'

Seems like there should be a data.py somewhere? As well as denoisenet_full.py and config_refine.py?

Sorry if I’m missing something obvious – I’m still pretty new to all this stuff.

[yuguochencuc]

Thanks for your comments. In fact, we only release the code of network architecture (denoisenet_high.py and denoisenet_low.py) before. We will release the rest code soon. Thanks for your attention.

[ndisci]

Hi, Thanks for your work. When did you release the rest code ?

[ariel-c-verbit]

Hey, any plans to release the rest of the code + a pre-trained model?

[yuguochencuc]

Hey, any plans to release the rest of the code + a pre-trained model?

Hey, actually，we have released all the code for training and enhancing. The pre-trained model for VB dataset is also released.

[yuguochencuc]

Hey, actually，we have released all the code for training and enhancing. The pre-trained model for VB dataset is also released.

[ndisci]

Hi, Is the released pre-trained model best model for your work ? According to paper you got 3.02 pesq score but when i tried on the same dataset the result was not cool. Thanks!

[yuguochencuc]

Hi, Is the released pre-trained model best model for your work ? According to paper you got 3.02 pesq score but when i tried on the same dataset the result was not cool. Thanks!

Hi, the pretrained model is not the best model in our paper (PESQ 3.02). In our paper, we use a two-stage training strategy, while the released pretrained model is trained on one-stage strategy (training the low-band and high-band together). So the performance is worse than the that on the two-stage training proceduce. We will release the best model soon.

[ndisci]

I got it. Thank you !