this didnt happen for me with the coco dataset itself because there were always class ids, but for another dataset it blew up.

* Remove dropout layer. Unnecessary.
* Reduce anchor stride from 2 to 1
* Increase training proposals to 200/image
* Reduce pre-NMS limit to 6000
* Improve computing proposal post:neg ratio
* Updated COCO training schedule

Reduces odds of weight explosion.

function cv2.rectangle, cv2.circle and cv2.fillPoly returns None.