YOLOv8 using TensorRT accelerate !

1. Install CUDA follow CUDA official website.

   🚀 RECOMMENDED CUDA >= 11.4

2. Install TensorRT follow TensorRT official website.

   🚀 RECOMMENDED TensorRT >= 8.4

3. Install python requirements.

   pip install -r requirements.txt

4. Install ultralytics package for ONNX export or TensorRT API building.

   pip install ultralytics

5. Prepare your own PyTorch weight such as yolov8s.pt or yolov8s-seg.pt.

NOTICE:

Please use the latest CUDA and TensorRT, so that you can achieve the fastest speed !

If you have to use a lower version of CUDA and TensorRT, please read the relevant issues carefully !

If you get ONNX from origin ultralytics repo, you should build engine by yourself.

You can only use the c++ inference code to deserialize the engine and do inference.

You can get more information in Normal.md !

Besides, other scripts won't work.

You can export your onnx model by ultralytics API and add postprocess such as bbox decoder and NMS into ONNX model at the same time.

python3 export-det.py \
--weights yolov8s.pt \
--iou-thres 0.65 \
--conf-thres 0.25 \
--topk 100 \
--opset 11 \
--sim \
--input-shape 1 3 640 640 \
--device cuda:0

• --weights : The PyTorch model you trained.
• --iou-thres : IOU threshold for NMS plugin.
• --conf-thres : Confidence threshold for NMS plugin.
• --topk : Max number of detection bboxes.
• --opset : ONNX opset version, default is 11.
• --sim : Whether to simplify your onnx model.
• --input-shape : Input shape for you model, should be 4 dimensions.
• --device : The CUDA deivce you export engine .

You will get an onnx model whose prefix is the same as input weights.

You can export TensorRT engine from ONNX by build.py .

Usage:

python3 build.py \
--weights yolov8s.onnx \
--iou-thres 0.65 \
--conf-thres 0.25 \
--topk 100 \
--fp16  \
--device cuda:0

• --weights : The ONNX model you download.
• --iou-thres : IOU threshold for NMS plugin.
• --conf-thres : Confidence threshold for NMS plugin.
• --topk : Max number of detection bboxes.
• --fp16 : Whether to export half-precision engine.
• --device : The CUDA deivce you export engine .

You can modify iou-thres conf-thres topk by yourself.

You can export TensorRT engine by trtexec tools.

Usage:

/usr/src/tensorrt/bin/trtexec \
--onnx=yolov8s.onnx \
--saveEngine=yolov8s.engine \
--fp16

If you installed TensorRT by a debian package, then the installation path of trtexec is /usr/src/tensorrt/bin/trtexec

If you installed TensorRT by a tar package, then the installation path of trtexec is under the bin folder in the path you decompressed

Please see more information in API-Build.md

Notice !!! We don't support YOLOv8-seg model now !!!

You can infer images with the engine by infer-det.py .

Usage:

python3 infer-det.py \
--engine yolov8s.engine \
--imgs data \
--show \
--out-dir outputs \
--device cuda:0

• --engine : The Engine you export.
• --imgs : The images path you want to detect.
• --show : Whether to show detection results.
• --out-dir : Where to save detection results images. It will not work when use --show flag.
• --device : The CUDA deivce you use.
• --profile : Profile the TensorRT engine.

You can infer with c++ in csrc/detect/end2end .

Please set you own librarys in CMakeLists.txt and modify CLASS_NAMES and COLORS in main.cpp.

export root=${PWD}
cd csrc/detect/end2end
mkdir -p build && cd build
cmake ..
make
mv yolov8 ${root}
cd ${root}

Usage:

# infer image
./yolov8 yolov8s.engine data/bus.jpg
# infer images
./yolov8 yolov8s.engine data
# infer video
./yolov8 yolov8s.engine data/test.mp4 # the video path

Please see more information in Segment.md

Please see more information in Pose.md

Please see more information in Cls.md

Please see more information in Obb.md

See more in README.md

Only test on Jetson-NX 4GB. See more in Jetson.md

If you want to profile the TensorRT engine:

Usage:

python3 trt-profile.py --engine yolov8s.engine --device cuda:0

If you need to break away from pytorch and use tensorrt inference, you can get more information in infer-det-without-torch.py, the usage is the same as the pytorch version, but its performance is much worse.

You can use cuda-python or pycuda for inference. Please install by such command:

pip install cuda-python
# or
pip install pycuda

Usage:

python3 infer-det-without-torch.py \
--engine yolov8s.engine \
--imgs data \
--show \
--out-dir outputs \
--method cudart

• --engine : The Engine you export.
• --imgs : The images path you want to detect.
• --show : Whether to show detection results.
• --out-dir : Where to save detection results images. It will not work when use --show flag.
• --method : Choose cudart or pycuda, default is cudart.