yolov5 yolov8 yolov9 speed test on T4 (tensorrt )

ausk commented 5 months ago

yolov5 yolov8 yolov9 speed test on T4 (tensorrt )

platform : T4; c++ with nms ; nchw=(1,3,384,640); input (1920, 1080); 200 loops, (already warmed up)

id	model	fps	size	map (paper)
01	yolov5s-v6.1	fps = 398.41	size = 15.14 MB	37.4
02	yolov5x-v6.1	fps = 101.42	size = 168.02 MB	50.7
03	yolov8n-v8.1	fps = 404.86	size = 7.44 MB	37.3
04	yolov8s-v8.1	fps = 365.63	size = 22.57 MB	44.9
05	yolov8m-v8.1	fps = 202.63	size = 51.15 MB	50.2
06	yolov8l-v8.1	fps = 136.71	size = 85.21 MB	52.9
07	yolov8x-v8.1	fps = 82.64	size = 132.23 MB	53.9
08	yolov9-c-converted	fps = 154.20	size = 50.99 MB	53.0
09	yolov9-e-converted	fps = 73.45	size = 114.07 MB	55.6

Based on the data, yolov9-c-converted outperforms yolov8l-v8.1 and yolov5x-v6.1 in terms of both fps and model size. Additionally, yolov9-c-converted also slightly surpasses yolov8l-v8.1 and yolov5x-v6.1 in accuracy(MAP).

从数据上看，yolov9-c-converted在帧率和模型大小方面都表现优于yolov8l-v8.1和yolov5x-v6.1。同时，yolov9-c-converted在精度上也略高于yolov8l-v8.1, 高于 yolov5x-v6.1

WongKinYiu commented 5 months ago

You may use wrong scripts to convert yolov9 models to tensorrt. If your conversion is correct, size of yolov9-c-converted will be about 52% size of yolov8l-v8.1.

We have tested in multiple methods include with/without nms. yolov9-c-converted is about 15~25% faster than yolov8l-v8.1. I suggest you to use official yolov8 code, or tensorrt-yolo to examine speed. We have checked above implementation.

ausk commented 5 months ago

@WongKinYiu I check, and find the yolov9 trt's type is fp32, while others type is fp16. I'll update the table soon.

ausk commented 5 months ago

@WongKinYiu Updated, all of them are the type of fp16 now.

Based on the data, yolov9-c-converted outperforms yolov8l-v8.1 and yolov5x-v6.1 in terms of both fps and model size. Additionally, yolov9-c-converted also slightly surpasses yolov8l-v8.1 and yolov5x-v6.1 in accuracy(MAP).

Thank you for such a great job.

WongKinYiu commented 5 months ago

Thanks.