#Normalizing Components of an Adversarial VAE's Loss Function

4 messages · Page 1 of 1 (latest)

hushed mist
#

Hey all! I need a sanity check.

MSE has a maximum possible value, and a discriminator can be made to output probabilities of real vs fake between 0 and 1. This should normalize the two values such that all I need to control is their relative weights as hyper parameters.

Sane?

west sigil
#

it has min value but max value is not limited

hushed mist
#

The max MSE value for any pixel is 1 * channels. So the max MSE value for an image is img_height * img_width * channels