use quantized or not? about fq-gan HOT 1 CLOSED

Nice question! The loss term acts as an implicit feature matching regularizer as demonstrated in the paper. In practice, we find that 1) the GAN training can benefit from this loss term; 2) When the quantized is returned and inputs to the next layer, parameters of the dictionary need an elaborate tuning. A warmup stage is sometimes designed to make sure the dictionary can catch up with the GAN training. This version's code is to make sure everyone can get expected and fancy GAN models without tuning anything.

from fq-gan.