TensorRT-YOLOv4

demo

Performance

model	input_size	GPU	mode	inference Time
yolov4	608x608	gtx 1080Ti	float32	23.3 ms
yolov4	416x416	gtx 1080Ti	float32	13.0 ms
yolov3	608x608	gtx 1080Ti	float32	18.2 ms
yolov3	416x416	gtx 1080Ti	float32	10.0 ms
yolov3-tiny	608x608	gtx 1080Ti	float32	3.31 ms
yolov3-tiny	416x416	gtx 1080Ti	float32	2.01 ms
yolov3-tiny-prn	608x608	gtx 1080Ti	float32	3.05 ms
yolov3-tiny-prn	416x416	gtx 1080Ti	float32	2.01 ms

1. Including pre-processing and post-processing time.

Enviroments

1. gtx 1080Ti

ubuntu 1604
TensorRT 5.0
cuda 9.0
python3 onnx=1.4.1 

Models

1. Add (infer_thresh) and (down_stride) to your .cfg.

[yolo]  ## small anchor
mask = 0,1,2
anchors = ...
down_stride = 8
infer_thresh = 0.5  
...
[yolo] ## mid anchor
mask = 3,4,5
anchors = ...
down_stride = 16
infer_thresh = 0.5
...
[yolo] ## big anchor
mask = 6,7,8
anchors = .....
down_stride = 32
infer_thresh = 0.5

1. Convert darknet yolo to onnx.

python3 tools/yolo_to_onnx.py --cfg model/yolov4.cfg --weights model/yolov4.weights --out model/yolov4.onnx

Example

git clone https://github.com/CaoWGG/TensorRT-YOLOv4.git
cd TensorRT-YOLOv4
mkdir build
cd build && cmake .. && make
cd ..
## yolov3
./buildEngine -i model/yolov3.onnx -o model/yolov3.engine -b 1 -m 0
./runDet -i model/yolov3.engine -p dog.jpg -v nuscenes_mini.mp4
## yolov4
./buildEngine -i model/yolov4.onnx -o model/yolov4.engine -b 1 -m 0
./runDet -i model/yolov4.engine -p dog.jpg -v nuscenes_mini.mp4

Related projects

• DarkNet
• TensorRT-CenterNet
• TensorRT
• onnx-tensorrt