search

aitorzip / PyTorch-SRGAN

A modern PyTorch implementation of SRGAN
https://arxiv.org/abs/1609.04802
GNU General Public License v3.0
361 stars 92 forks source link

high resolution images of dimension 96X96? #11

Open udion opened 6 years ago

udion commented 6 years ago

How did you get 96X96 high resolution REAL images for comparison? My understanding was that we take the original dataset (cifar-10, consisting images of dimension 32X32), down-sample the images to (say 4 times, i.e 8X8) and use this down-sampled and corresponding original version for the training pair.

While testing we will feed some down-sampled (i.e 8X8) and expect a 32X32 image which is very close to corresponding original 32X32 image. How come your outputs are 96X96? It seems you first up-scaled the images and then down-sampled. Will it not affect the quality of the output?

aitorzip commented 6 years ago

Hi, the results for the 96x96 resolution were trained using ImageNet, not CIFAR-10. Doing it as you say would be silly as you correctly guess.

udion commented 6 years ago

also did you tune the beta in swish activation function?