Yolor-ROS

A set of python scripts implementing Yolor with TensorRT for object detection in ROS workspace.

Preparation

1. Clone this resipiratory to your catkin workspace

cd /catkin_ws/src
git clone https://github.com/petpetpeter/Yolor-ROS.git

2. Build your workspace

cd /catkin_ws/
catkin build
source ~/.bashrc

3. Install Python Dependencies

install pytorch according to your environment https://pytorch.org/

roscd yoloros/scripts/yolor
pip install -e .

4. Download pretrain weight to /scripts/yolor

https://drive.google.com/file/d/1_vgbLEjV0TU5492MfNEkJduTdGtcdthN/view?usp=sharing

5. Launch ROS Yolor example

roslaunch yoloros yolor_camera.launch

TensorRT 8.0.3.4: Follow installation guide here https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html

1. Convert pytorch weight file to onnx file

roscd yoloros/scripts/yolor
python convert2Onnx.py --weight weights/yolor_p6.pt --out onnx/yolor_p6.onnx 

2. Convert onnx to tensorRT Engine

/usr/src/tensorrt/bin/trtexec --explicitBatch \
                                --onnx=transformer_encoder.onnx \
                                --saveEngine=transformer_encoder.trt \
                                --minShapes=input:1x3x32x128 \
                                --optShapes=input:32x3x32x512 \
                                --maxShapes=input:32x3x32x768 \
                                --verbose \
                                --fp16

3. Launch ROS Yolor TensorRT example

roslaunch yoloros trt_camera.launch

Acknowledgement

• Yolor: https://github.com/WongKinYiu/yolor
• Yolor-TensorRT: https://github.com/NNDam/yolor