Great chart.
Dev Nag

Oh, thanks awesome tips. Indeed it (somehow) gets more steady with a bit of tweaking :-) (specifically towards the end; though it would be the same if we don’t tweak but stop training a bit earlier ~23k epoch).

It looks like the D/G errors and their output are somehow analogous.

Before tweaking:

GANs errors and output 1 — PyTorch


GANs errors and output 2 — PyTorch