Prediction is just white area when using pconv mask

[qwerdbeta]

Hello, I get good results when training with square mask. But when I use pconv masks images (black and white where white is masked area), all predictions are just white pixels. No shapes or colors produced in pconv mode. tried using hinge gan type for both. Also tried nsgan for a bit with no luck.

[qwerdbeta]

I gave up on pconv because I really just needed a on square rectangle which I achieved. However, I'm wondering what weights in the Json files might be better tuned for 522x512 images (larger) plus larger masks and simple images. I noticed did used nsgan instead of hinge gan type

[ZHANG-hengzhi]

你好，请问怎么构建自己的数据集呀，config里该如何设置自己的路径，可以分享一下吗