发帖
雷达卡
Improved Contrastive Divergence Training of Energy-Based Models
Yilun Du 1 Shuang Li 1 Joshua Tenenbaum 1 Igor Mordatch 2
Abstract
Contrastive divergence is a popular method of
training energy-based models, but is known to
have difficulties with training stability. We pro-
pose an adaptation to improve contrastive diver-
gence training by scrutinizing a gradient term that Figure 1: (Left) 128x128 samples on unconditional CelebA-HQ.
...
京公网安备 11010802022788号
