The snap package is automatically built and released to the snap store on each commit to the main branch. It is currently built for amd64 and arm64 platforms. Unfortunately, it will probably only work with CPU and not Cuda because of difficulties packaging Cuda libraries in the snap.
Install the stable version:
sudo snap install easyvizar-detect
Install the latest build:
sudo snap install --edge easyvizar-detect
The install will take some time to complete, as it downloads and installs PyTorch as part of the install operation.
If this will be running on the same machine as the easyvizar-edge snap (recommended), you should connect the easyvizar-edge:data interface to allow easyvizar-detect to read image files stored by easyvizar-edge.
sudo snap connect easyvizar-detect:data easyvizar-edge:data
In order to take advantage of the GPU, it is recommended to run the Python code directly on the Jetson board without Docker or Snap confinement. Start by using the Nvidia SDK tools to set up the required drivers.
Find a version of the PyTorch wheel that is compatible with your Jetson. https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-11-now-available/72048
Find a version of torchvision that is compatible. https://pypi.org/project/torchvision/ https://qengineering.eu/install-pytorch-on-jetson-nano.html
Be careful to use a compatible version of Python that is compatible with the dependencies as well as Nvidia's PyTorch wheel. Python 3.6 with torch 1.10.0 and torchvision 0.11.0 seems to work.
sudo apt-get install libopenblas-base libopenmpi-dev
python3.6 -m pip install --upgrade pip
python3.6 -m pip install -r requirements.txt
python3.6 -m pip install torch-1.10.0-cp36-cp36m-linux_aarch64.whl
python3.6 -m pip install torchvision-0.11.0a0+fa347eb-cp36-cp36m-linux_aarch64.whl
If everything is set up correctly, the following command should print "True".
python3.6 -c "import torch; print(torch.cuda.is_available())"
Finally, run the detector:
python3.6 -m detect
docker build -t detect
docker run -d --gpus all --add-host=host.docker.internal:host-gateway -e VIZAR_SERVER="http://host.docker.internal:5000" --restart always --name detect detect
Download a pretrained model file.
https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8m-seg.pt
Use the yolo export script.
yolo export model=yolov8m-seg.pt format=onnx imgsz=428,760 opset=16
Use the augment script to add an NMS layer.
python3 scripts/augment_model.py yolov8m-seg.onnx
Use netron to verify that the model looks correct.
netron yolov8m-seg-nms.onnx