Number of Classes for Faster RCNN Multi-Class Object Detection

I have a dictionary of classes numbered from 1-19 (starts with 1).
What should be the num_classes for the Faster RCNN predictor?
Should it be 20 (num of classes + 1) or still just 19?

Generally the class indices start at 0, so if you are dealing with 19 classes, you would use indices in [0, 18] unless you want to reserve class index 0 for a specific class such as the background class.