Repository navigation
Some of the results are full of noise. #93
Description
Activity
I had the same issue when I adopted the colorization data loader to my own MR Image translation task. When I removed the normalization in the transforms, the noise was gone. But the psnr and ssim were very poor.
I had the same issue when I adopted the colorization data loader to my own MR Image translation task. When I removed the normalization in the transforms, the noise was gone. But the psnr and ssim were very poor.
Ok, thanks for your answer. I'll have a try on it. Normalization maybe is connected by your data distribution.
@TMDTom @SarielMa Now my MRI translation task meets the same problem, around 80% of the outputs are successful, and the rest are very noisy. And this happens after I set the noise scheduler to "cosine", before that, I couldn't even get a single successful output. So did you solve the problem? And how? Thanks a lot!
@AwayGithub
I also have same issue.
I trained and tested my own dataset.
But my test results are some of them were good and some of them were just noise.
My dataset was 580 size of 256*256 and I use colorization method.
I check every 1000 epoch until 5000epochs. and It was getting better but still have randomly noise results.
Did you solve it?@AwayGithub I also have same issue. I trained and tested my own dataset. But my test results are some of them were good and some of them were just noise. My dataset was 580 size of 256*256 and I use colorization method. I check every 1000 epoch until 5000epochs. and It was getting better but still have randomly noise results. Did you solve it?
Maybe you can try the way: change the initial data distribution to 0-1 or -1-1.
Some of the results are full of noise. And there are few results successful to me. Maybe like this.
