D is much stronger than G when training VOC2012 causing gradient vanishing

[syfbme]

I have to set D's learning rate be 1/10 of G. I wonder if anyone meet my issue. In my training, the D convergences early after 2 or 3 epoches

In official code, the losses are added without multiple a lambda number. In my training result, the images are too blur. Do i need to add a big lambda on GAN Loss?

[WayneCho]

I think you can modify line 109 in pix2pixHD_model.py.