How can I retrain this model? about 2018_subeesh_epsr_eccvw HOT 6 CLOSED

Hello,

One can use the code released by authors of EDSR (https://github.com/thstkdgus35/EDSR-PyTorch) to reproduce the results. Run the code following the format in - (https://github.com/thstkdgus35/EDSR-PyTorch/blob/master/src/demo.sh). We used the following training options, --template GAN, --gan_k 2, --res_scale 0.1, --patch_size 192, and --loss p1 x VGG54+p2 x GAN+p3 x MSE. The exact parameter values used in our experiments can be found in our paper (https://arxiv.org/pdf/1811.00344.pdf).

from 2018_subeesh_epsr_eccvw.

Thanks!

from 2018_subeesh_epsr_eccvw.

Are the values of these parameters the same as the one you used for region 1 of PIRM2018? I am asking because these values do not match the default values of options.py in this repo.

from 2018_subeesh_epsr_eccvw.

Also, which pre-trained model file do I supply for weight initialization if I'm retraining EPSR (region 1) from the EDSR repo?

from 2018_subeesh_epsr_eccvw.

Are the values of these parameters the same as the one you used for region 1 of PIRM2018? I am asking because these values do not match the default values of options.py in this repo.

Yes. The parameter values of my experiments are mentioned in the paper

from 2018_subeesh_epsr_eccvw.

Also, which pre-trained model file do I supply for weight initialization if I'm retraining EPSR (region 1) from the EDSR repo?

I have mentioned it in the paper, "we initialized the model weights of generator using pre-trained weights of EDSR
(obtained by training EDSR with λ1 = λ3 = 0)." You can either use the pretrained weights supplied by authors of EDSR (this should be easy to find out), or else train EDSR with this settings by urself and then use those weights.

from 2018_subeesh_epsr_eccvw.