Incompatible changes about head design:

• Mask head and keypoint head now include logic for losses & inference. Custom heads should overwrite the feature computation by layers() method.
• _forward_{box,mask,keypoint} methods of StandardROIHeads now accept dict of features.

This release is made to be compatible with such changes in projects (Mesh R-CNN, PointRend, etc)

Other additional features:

• flops & parameters counting
• Improve speed & RAM consumption of data loader

The pre-built wheels for this version have to be used with an official binary release of PyTorch 1.4.