I tested on one video and pytorch without NMS gives a lot better FPS that TRT inference I have shared the code that I used to inference TRT and for PyTorch I used detect.py in ultralytics/yolov5 repo to compute FPS (Both calculated without NMS).
Image Size -> 640 * 640
PyTorch(without NMS) -> 125+
TRT (without NMS) -> 105
I tested on one video and pytorch without NMS gives a lot better FPS that TRT inference I have shared the code that I used to inference TRT and for PyTorch I used detect.py in ultralytics/yolov5 repo to compute FPS (Both calculated without NMS). Image Size -> 640 * 640 PyTorch(without NMS) -> 125+ TRT (without NMS) -> 105