search

xinntao / ESRGAN

ECCV18 Workshops - Enhanced SRGAN. Champion PIRM Challenge on Perceptual Super-Resolution. The training codes are in BasicSR.
https://github.com/xinntao/BasicSR
Apache License 2.0
5.86k stars 1.04k forks source link

Can you show an example of an SR image after training stage 1? #109

Closed qAp closed 3 years ago

qAp commented 3 years ago

In your paper you say that "The training process is divided into two stages. First, we train a PSNR-oriented model with the L1 loss. The learning rate is initialized as 2×10^−4 and decayed by a factor of 2 every 2×10^5 of mini-batch updates."

Can you please show an example of an image output by the generator, along with the real image, after this first stage is completed? It'd really help me know if I'm on the right track.

The number of mini-batches to train for in this stage 1. Is it just enough to get the PSNR to level off?

Slightly unrelated. The generation of low resolution images uses the 'bicubic' method. Does ESRGAN only work for this method, or the algorithm doesn't really depend on the way with which the LR images are generated?

Thanks

xinntao commented 3 years ago

We have released the first stage model: https://drive.google.com/drive/folders/15DgDtfaLASQ3iAPJEVHQF49g9msexECG for ESRGAN_PSNR_SRx4_DF2K_official-150ff491.pth

now ESRGAN relies heavily on the trained bicubic kernel.