pabsan-0/yolov8-orin

Sample workspace to quickly deploy yolo models on NVIDIA orin

NVIDIA Orin ultralytics

Sample workspace to quickly deploy yolo models on NVIDIA ORIN.

Dependencies

• Docker
• Jetpack 5.1.1
• Docker compose: apt install docker-compose

Usage

• Clone this repo git clone --recursive https://github.com/pabsan-0/yolov8-orin
• Convert model from yolov8 to darknet:

$ docker-compose run --rm ultralytics
# bash dl_weights.sh   # wget weights or get your own
# python3 gen_wts_yolov8.py --size 640 -w yolov8n.pt -o /weights
# rm labels.txt

• Run deepstream on the converted model (darknet->tensorrt is automatic):

$ cd DeepStream-Yolo
$ ls -l /usr/local/ | grep cuda  ## check cuda version
$ CUDA_VER=11.4 make -C nvdsinfer_custom_impl_Yolo # Compilation needs to be done on host
$ docker-compose run deepstream 
# bash pipe_sh/v4l2-docker.py  

The first time you convert an .engine, dump it in /weights to avoid rebuilds.

Snippets & links

• Check cuda version ls -l /usr/local | grep cuda
• Check deepstream version & health: deepstream-app --version
• Jetpack containers for jetson:
  • Pytorch: l4t-pytorch
  • Deepstream: deepstream-l4t

Bugs and limitations

• Yolov8 may need a few extra dependencies for its built-in engine export
  • pip3 install --no-cache onnx>=1.12.0 onnxsim>=0.4.17
  • Get and install a wheel for onnxruntime for your Jetpack at https://elinux.org/Jetson_Zoo#ONNX_Runtime
• Gstreamer pipelines in C do not allow ' characters
• Deepstream 6.2 sometimes requires specifying video format in pre-nvstreammux caps:

nvvideoconvert ! 'video/x-raw(memory:NVMM),format=NV12'