training problem！！！！！
#31
by yuduan - opened
Hello, I have a question. I followed the Selforcing Plus setup in section 2.2 (High Noise) and ran it with backward simulation. The timestep is set between 900–1000, shift = 5, denoising_step_list = [990, 960, 930, 900], minstep = 900, maxstep = 990, dfake = 5. I’m training on 16 GPUs. The DMD loss decreased from 0.8 to 0.3, while the critic loss went from 0.5 up to as high as 3, and is now oscillating around 1.5 after about 600 steps. Is this normal?
You re trying to fix it?