this article is very helpful to me thanks >//////<

BTW, in my case TensorRT only improves around 1.5~2 speed on YOLOv3, which doesn’t achieve my expectation QAQ (but is enough for my application)