Allow mask rcnn to be run with mixed precision without NaNs.
Some parts of the forward pass would previously overflow in float16. Such parts are now done in float32. PiperOrigin-RevId: 380857663
Showing
想要评论请 注册 或 登录
Some parts of the forward pass would previously overflow in float16. Such parts are now done in float32. PiperOrigin-RevId: 380857663