This project provides a detailed tutorial for how to use yolov7 in ROS based on the acceleration of tensorrt.
Please find the Chinese user guide ./yolov7_tensorrt_ros/使用教程.pdf.
The English user guide as follows:
yolov7 https://github.com/WongKinYiu/yolov7
YOLOv7_Tensorrt https://github.com/Monday-Leo/YOLOv7_Tensorrt
cd yolov7_tensorrt_ros
Create the conda environment
conda create -n name python=3.9
pip install -r requirements.txt
Use the following instructions to make sure that you have correctly install CUDA and CUDNN
NVCC -V
cat /user/local/cuda/include/cudnn_version.h | grep CUDNN_MAJOR -A 2
Download the tensorrt Click into https://developer.nvidia.com/nvidia-tensorrt-8x-download
Find the suitable version, download it and install it.
tar -xzvf TensorRT-8.4.1.5.Linux.x86_64-gnu.cuda-11.6.cudnn8.4.tar.gz
vim ~/.bashrc
put the following sentence at the bottom of the ./bashrc
# export LD_LIBRARY_PATH=$LD_LIBRARY_PATH：../TensorRT-8.4.1.5/lib
source ~/.bashrc
Run the samples to make sure that you have install the tensorrt successfully.
cd ../TensorRT-8.4.1.5/samples/sampleMNIST
make -j16
cd ../../bin/
./sample_mnist
If you find the results like
or
That means you have successfully install the tensorrt
sudo gedit /etc/ld.so.conf
../TensorRT-8.4.1.5/lib
sudo ldconfig
sudo ldconfig -v
sudo ln -sf /usr/local/cuda/targets/x86_64-linux/lib/libcudnn.so.8.4.0
/usr/local/cuda/targets/x86_64-linux/lib/libcudnn.so.8
sudo cp ../TensorRT-8.4.1.5/lib/libnvinfer_builder_resource.so.8.4.1
/usr/lib
git clone https://github.com/WongKinYiu/yolov7
git clone https://github.com/Monday-Leo/YOLOv7_Tensorrt
Add YOLOv7_ EfficientNMS.py and export of Tensorrt warehouse_ Copy onnx. py to yolov7 and export it with Efficient NMS ONNX model (remember to install the conda environment before opening it.
python export_onnx.py --weights ./weights/yolov7.pt
If you don't have the pre-trained model of yolo, download at https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7.pt and put it into ./weights(If you dont.t have please create one).
Copy the converted onnx model to the Tensorrt/bin folder, and convert it to the onnx model of Efficient NMS Type. FP32 prediction deletion - fp16 parameter is sufficient.
trtexec --onnx=./yolov7.onnx --saveEngine=./yolov7_fp16.engine --fp16 --workspace=200
After waiting for the serialization model to be formed, modify YOLOv7_ Tensorrt warehouse infer.py model path and diagram path.
trt_engine = TRT_engine("./trt_model/yolov7_fp16.engine")
img1 = cv2.imread("./pictures/zidane.jpg")
Run the test example
python infer.py
roslaunch kinect2_bridge kinect2_bridge.launch
rosrun yolov7_tensorrt_ros yolo_intermediate.py
