search

visinf / n3net

Neural Nearest Neighbors Networks (NIPS*2018)
Other
287 stars 45 forks source link

Training details for real image denoising #4

Closed Shakarim94 closed 5 years ago

Shakarim94 commented 6 years ago

BSDS, DIV2K and Waterloo datasets are used in the paper for real image denoising. I want to know how many patches were extracted for training. Did you use the same principle as in Gaussian denoising (512 80x80 patches from each image)?

What about the number of epochs and lr decay? I couldn't find those on the paper, they are given only for Gaussian denoising.

tobiasploetz commented 6 years ago

There are 50 epochs, with a decay as specified in here. The initial learning rate is 0.001. Each epoch sweeps 16 times through the training set, cropping a 80x80 patch from each image.

I also tried a longer training (more iterations per epoch) but that did not result in better accuracy.