imgcomp-cvpr
imgcomp-cvpr copied to clipboard
about soft and hard quantization
hello, I don't figure out that when testing the model , why to use the hardout for the input of the P network .In my opinions ,the hardout and the symbols are equal in a sense .
I think one contains values at the symbols. e.g. if your quantization grid is:
levels = [0.2, 0.5, 0.7]
and your symbols are
sym = [0, 0, 1, 0, 2]
then we would ahve
hardout = [0.2, 0.2, 0.5, 0.2, 0.7]
The important bit being that we can have non-uniform quantization.
Thanks a lot, I have figured it out.^_^