steven
steven
> The artifacts are likely to be caused by unrobustness of LR generation for some unseen distributions. One solution is also to finetune models on several samples of the target...
> In config files, e.g. options/train/train_IRN_x4.yml, modify the variable "lambda_fit_forw", e.g. from 16 to 160. Much appreciated!
I've continually trained the pretrained model 'IRN_x2.pth' with another 2500 epochs with kodak image set.As a result, the kodak images' artifacts are gone, but the DIV2k images have artifacts however....
After I changed '**the loss weight for LR guidance(from 4 to 40)**' and continually trained the pretrained 'IRN_x2' model with 5000 epochs, the model is more robust with all these...
> > After I changed '**the loss weight for LR guidance(from 4 to 40)**' and continually trained the pretrained 'IRN_x2' model with 5000 epochs, the model is more robust with...
> > > > After I changed '**the loss weight for LR guidance(from 4 to 40)**' and continually trained the pretrained 'IRN_x2' model with 5000 epochs, the model is more...
> > > > > > After I changed '**the loss weight for LR guidance(from 4 to 40)**' and continually trained the pretrained 'IRN_x2' model with 5000 epochs, the model...
> > > > > > > > After I changed '**the loss weight for LR guidance(from 4 to 40)**' and continually trained the pretrained 'IRN_x2' model with 5000 epochs,...
> Use 'DIV2k' too.
> > > > > > > > > Use 'DIV2k' too. > > you mean the niter = 5000? Yes!Sorry for misleading you.