Summary:
Make training much more stable (no NaNs in 7 runs).
Results slightly better (average is around 78.8).

Reviewed By: alexander-kirillov

Differential Revision: D22427114

fbshipit-source-id: 3a3ea9895e5ac955a62454b73cff71f496e0fbfb

Summary: Pull Request resolved: https://github.com/fairinternal/detectron2/pull/431

Reviewed By: alexander-kirillov

Differential Revision: D22266546

fbshipit-source-id: c1d585d66edf0eb85e84ba0b39bca338ab892b6b

Summary: semantic segmentation PointRend

Reviewed By: ppwwyyxx

Differential Revision: D19350389

fbshipit-source-id: cec04422ae5b76d730257de336d871cf281b625a