How to modify region proposal network proposals to speed up inference

I am using the config traffic_lights_spec_resnet10_sample_image.txt (3.5 KB)

especially the part :
rpn_pre_nms_top_N: 6000
rpn_nms_max_boxes: 300
rpn_nms_overlap_threshold: 0.2
bbox_visualize_threshold: 0.2
classifier_nms_max_boxes: 300
classifier_nms_overlap_threshold: 0.3

As I am trying to detect really small objects in a large image I wanted to reduce the praposals to around 50 or so to speed up inference on nvidia xavier

Please consider below.

rpn_pre_nms_top_N: 1000
rpn_nms_max_boxes: 50
classifier_nms_max_boxes: 50

I will have to retrain the network or make the change during evaluation ?

No need to retrain because it is the hyper-parameter in evaluation config.