Losses

[woctezuma]

I think it is not clear what the losses displayed in the log are.

lightweight-gan/lightweight_gan/lightweight_gan.py

Lines 1184 to 1191 in 9a48e87

	def print_log(self):
	data = [
	('G', self.g_loss),
	('D', self.d_loss),
	('GP', self.last_gp_loss),
	('SS', self.last_recon_loss),
	('FID', self.last_fid)
	]

My understanding is that, in G: 1.37 | D: 1.80 | GP: 0.97 | SS: 0.02, we have:

• the loss $L_G$ for the Generator (G),
• the (total) loss $L_D$ for the Discriminator (D),
• a loss from Gradient Penalty (GP),
• a reconstruction loss from the Self-Supervision (SS) of the the discriminator.

It would be nice to mention it somewhere so that we can try to understand what happens during training. :)

[fabiooshiro]

A noob question: what is a good/bad result range?

my output

G: 2.32 | D: 0.04 | GP: 0.09 | SS: 0.00
G: 1.93 | D: 0.10 | GP: 0.03 | SS: 0.00
G: 2.52 | D: 0.14 | GP: 0.06 | SS: 0.00
G: 2.45 | D: 0.09 | GP: 0.13 | SS: 0.00
G: 1.32 | D: 0.09 | GP: 0.31 | SS: 0.00
G: 1.95 | D: 0.35 | GP: 0.15 | SS: 0.01
G: 2.10 | D: 0.10 | GP: 0.03 | SS: 0.00
G: 2.60 | D: 0.12 | GP: 0.08 | SS: 0.00
G: 1.99 | D: 0.87 | GP: 0.09 | SS: 0.00
G: 1.86 | D: 0.72 | GP: 0.10 | SS: 0.00