Comment by rspeer

There may be a way to build it into the loss function so that it happens before the quantization, right?

(and holy crap, look how fast the HN conservatives are getting to my comment)