How should I run the inference?

[AlpTuncay]

In this implementation of Faster RCNN, to my understanding, the inference in test_frcnn.py is run using two different networks: first, image is fed to network that consists of shared layers and RPN, second, the ROI output of the RPN is fed to classifier network. So there are two different neural networks and in between there are some processes done on the output of the first network. In the paper, however, it is stated that the system is a single, unified network. From this statement, I understand that when we want to run inference, there should be just one network in which the images are fed with no disconnection between two graphs and we get the bounding boxes and scores as a result.

1. Can someone tell me whether my understanding of the original paper is correct?
2. If so, is it possible to run inference as I explained above?