Open Ksky001 opened 4 years ago
The decoder of the generator takes a latent space tensor and a conditional attribute as its inputs. The attributes in CelebA dataset are 0 and 1. But we surmise that it would be better to have symmetric inputs like what we usually do to the images for GANs. We always normalize an images to (-1, 1) when doing the generative stuffs.
This line att_a_ = (att_a * 2 - 1) * thres_int
has the same effect. The thres_int
is the threshold intensity playing a role of a scaling factor. It is 0.5 by default. So the attributes, which are 0s and 1s, are hence normalized as (-0.5, 0.5).
Please let me know if you have any other questions.
why use extra attra and attrb for attr_a and attr_b?