Forked from https://github.com/rbgirshick/py-faster-rcnn
- Train-net has these elements (not in test-net)
- LR parameters, e.g. lr_muti
- Weight initializer, e.g. Gaussian mean = 0, std = 0.001
- Drop-out layers
- Test-net has RPN as part of the net model
- which train-net doesn't have