NOTE: I am doing post processing for restoration after CNN output. Lets say CNN output is A. Then final output is Final_OUTPUT = A + B.
During restoration of image, Final_OUTPUT values range goes beyond -1 to 1. When converted to image it looks like noise in image where the value exceed to less than -1 and greater than 1. Is it better to
(1) normalize the data ( x/torch.max(mod(x)) OR
(2) clip the less than -1 value to -1 and greater than 1 value to 1.
Which will be better when loss is backpropogate? Can CNN learn after many epoch to range the data between -1 to 1 in case of (1). I tried (2), but color changes as value clip to -1 and 1.