thangkt/PCB-Prune-YOLO-TensorRT-FP16

PCB-Prune-YOLO TensorRT FP16

Static batch-1 TensorRT FP16 engines for the baseline and direct DepGraph P10, P20, and P30 checkpoints. Engines were built and benchmarked directly on Tesla T4 with TensorRT 10.16.1.11, CUDA 12.8, PyTorch 2.10.0+cu128, and Ultralytics 8.4.115. Input is [1,3,640,640]; raw output is [1,10,8400]; NMS is external.

Model	Validation mAP50-95	Mean latency	FPS	Engine size
Baseline	0.78716	1.837 ms	544.28	7.477 MiB
P10 direct	0.77842	2.023 ms	494.34	7.627 MiB
P20 direct	0.76931	1.933 ms	517.45	7.378 MiB
P30 direct	0.75610	1.754 ms	569.97	5.482 MiB

Latency is pure synchronized engine forward after 50 warm-ups across 200 measurements and excludes preprocessing/NMS. TensorRT engines are not portable across arbitrary TensorRT/CUDA/GPU combinations; rebuild them on the deployment target when the stack differs. Full provenance and reports are included.

Project: https://github.com/pnthang04/PCB-Prune-YOLO