KernelGAN
KernelGAN copied to clipboard
Question about the DownScaleLoss
There is no DownScaleLoss in the paper, could we set the param 'lambda_bicubic' to 0?
It is referred to in sec 4.1:
To provide a reasonable initial starting point, the generator’s output is constrained to be similar to an ideal downscaling (e.g. bicubic) of the input, for the first iterations. Once satisfied, this constraint is discarded.
Note that lambda_bicubic
is decreased to 0 relatively fast