Quick Start Guide: NVIDIA Jetson with Ultralytics YOLOv8

This comprehensive guide provides a detailed walkthrough for deploying Ultralytics YOLOv8 on NVIDIA Jetson devices. Additionally, it showcases performance benchmarks to demonstrate the capabilities of YOLOv8 on these small and powerful devices.

What is NVIDIA Jetson?

NVIDIA Jetson is a series of embedded computing boards designed to bring accelerated AI (artificial intelligence) computing to edge devices. These compact and powerful devices are built around NVIDIA's GPU architecture and are capable of running complex AI algorithms and deep learning models directly on the device, without needing to rely on cloud computing resources. Jetson boards are often used in robotics, autonomous vehicles, industrial automation, and other applications where AI inference needs to be performed locally with low latency and high efficiency. Additionally, these boards are based on the ARM64 architecture and runs on lower power compared to traditional GPU computing devices.

NVIDIA Jetson Series Comparison

Jetson Orin is the latest iteration of the NVIDIA Jetson family based on NVIDIA Ampere architecture which brings drastically improved AI performance when compared to the previous generations. Below table compared few of the Jetson devices in the ecosystem.

For a more detailed comparison table, please visit the Technical Specifications section of official NVIDIA Jetson page.

What is NVIDIA JetPack?

NVIDIA JetPack SDK powering the Jetson modules is the most comprehensive solution and provides full development environment for building end-to-end accelerated AI applications and shortens time to market. JetPack includes Jetson Linux with bootloader, Linux kernel, Ubuntu desktop environment, and a complete set of libraries for acceleration of GPU computing, multimedia, graphics, and computer vision. It also includes samples, documentation, and developer tools for both host computer and developer kit, and supports higher level SDKs such as DeepStream for streaming video analytics, Isaac for robotics, and Riva for conversational AI.

Flash JetPack to NVIDIA Jetson

The first step after getting your hands on an NVIDIA Jetson device is to flash NVIDIA JetPack to the device. There are several different way of flashing NVIDIA Jetson devices.

1. If you own an official NVIDIA Development Kit such as the Jetson Orin Nano Developer Kit, you can visit this link to download an image and prepare an SD card with JetPack for booting the device.
2. If you own any other NVIDIA Development Kit, you can visit this link to flash JetPack to the device using SDK Manager.
3. If you own a Seeed Studio reComputer J4012 device, you can visit this link to flash JetPack to the included SSD.
4. If you own any other third party device powered by the NVIDIA Jetson module, it is recommended to follow command-line flashing by visiting this link.

Set Up Ultralytics

There are two ways of setting up Ultralytics package on NVIDIA Jetson to build your next Computer Vision project. You can use either of them.

• Start with Docker
• Start without Docker

Start with Docker

The fastest way to get started with Ultralytics YOLOv8 on NVIDIA Jetson is to run with pre-built docker image for Jetson.

Execute the below command to pull the Docker container and run on Jetson. This is based on l4t-pytorch docker image which contains PyTorch and Torchvision in a Python3 environment.

t=ultralytics/ultralytics:latest-jetson && sudo docker pull $t && sudo docker run -it --ipc=host --runtime=nvidia $t

After this is done, skip to Use TensorRT on NVIDIA Jetson section.

Start without Docker

Install Ultralytics Package

Here we will install ultralyics package on the Jetson with optional dependencies so that we can export the PyTorch models to other different formats. We will mainly focus on NVIDIA TensorRT exports because TensoRT will make sure we can get the maximum performance out of the Jetson devices.

1. Update packages list, install pip and upgrade to latest

   sudo apt update
   sudo apt install python3-pip -y
   pip install -U pip
   

2. Install ultralytics pip package with optional dependencies

   pip install ultralytics[export]
   

3. Reboot the device

   sudo reboot
   

Install PyTorch and Torchvision

The above ultralytics installation will install Torch and Torchvision. However, these 2 packages installed via pip are not compatible to run on Jetson platform which is based on ARM64 architecture. Therefore, we need to manually install pre-built PyTorch pip wheel and compile/ install Torchvision from source.

1. Uninstall currently installed PyTorch and Torchvision

   pip uninstall torch torchvision
   

2. Install PyTorch 2.1.0 according to JP5.1.3

   sudo apt-get install -y libopenblas-base libopenmpi-dev
   wget https://developer.download.nvidia.com/compute/redist/jp/v512/pytorch/torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl -O torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl
   pip install torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl
   

3. Install Torchvision v0.16.2 according to PyTorch v2.1.0

   sudo apt install -y libjpeg-dev zlib1g-dev
   git clone https://github.com/pytorch/vision torchvision
   cd torchvision
   git checkout v0.16.2
   python3 setup.py install --user
   

Visit the PyTorch for Jetson page to access all different versions of PyTorch for different JetPack versions. For a more detailed list on the PyTorch, Torchvision compatibility, visit the PyTorch and Torchvision compatibility page.

Install onnxruntime-gpu

The onnxruntime-gpu package hosted in PyPI does not have aarch64 binaries for the Jetson. So we need to manually install this package. This package is needed for some of the exports.

All different onnxruntime-gpu packages corresponsing to different JetPack and Python versions are listed here. However, here we will download and install onnxruntime-gpu 1.17.0 with Python3.8 support for the JetPack we are using for this guide.

wget https://nvidia.box.com/shared/static/zostg6agm00fb6t5uisw51qi6kpcuwzd.whl -O onnxruntime_gpu-1.17.0-cp38-cp38-linux_aarch64.whl
pip install onnxruntime_gpu-1.17.0-cp38-cp38-linux_aarch64.whl

Use TensorRT on NVIDIA Jetson

Out of all the model export formats supported by Ultralytics, TensorRT delivers the best inference performance when working with NVIDIA Jetson devices and our recommendation is to use TensorRT with Jetson. We also have a detailed document on TensorRT here.

Convert Model to TensorRT and Run Inference

The YOLOv8n model in PyTorch format is converted to TensorRT to run inference with the exported model.

from ultralytics import YOLO

# Load a YOLOv8n PyTorch model
model = YOLO('yolov8n.pt')

# Export the model
model.export(format='engine')  # creates 'yolov8n.engine'

# Load the exported TensorRT model
trt_model = YOLO('yolov8n.engine')

# Run inference
results = trt_model('https://ultralytics.com/images/bus.jpg')

# Export a YOLOv8n PyTorch model to TensorRT format
yolo export model=yolov8n.pt format=engine  # creates 'yolov8n.engine'

# Run inference with the exported model
yolo predict model=yolov8n.engine source='https://ultralytics.com/images/bus.jpg'

NVIDIA Jetson Orin YOLOv8 Benchmarks

YOLOv8 benchmarks were run by the Ultralytics team on 10 different model formats measuring speed and accuracy: PyTorch, TorchScript, ONNX, OpenVINO, TensorRT, TF SavedModel, TF Graphdef, TF Lite, PaddlePaddle, NCNN. Benchmarks were run on Seeed Studio reComputer J4012 powered by Jetson Orin NX 16GB device at FP32 precision with default input image size of 640.

Comparison Chart

Eventhough all model exports are working with NVIDIA Jetson, we have only included PyTorch, TorchScript, TensorRT for the comparison chart below because, they make use of the GPU on the Jetson and are guaranteed to produce the best results. All the other exports only utilize the CPU and the performance is not as good as the above three. You can find benchmarks for all exports in the section after this chart.

Detailed Comparison Table

The below table represents the benchmark results for five different models (YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, YOLOv8x) across ten different formats (PyTorch, TorchScript, ONNX, OpenVINO, TensorRT, TF SavedModel, TF Graphdef, TF Lite, PaddlePaddle, NCNN), giving us the status, size, mAP50-95(B) metric, and inference time for each combination.

Visit this link to explore more benchmarking efforts by Seeed Studio running on different versions of NVIDIA Jetson hardware.

Reproduce Our Results

To reproduce the above Ultralytics benchmarks on all export formats run this code:

from ultralytics import YOLO

# Load a YOLOv8n PyTorch model
model = YOLO('yolov8n.pt')

# Benchmark YOLOv8n speed and accuracy on the COCO8 dataset for all all export formats
results = model.benchmarks(data='coco8.yaml', imgsz=640)

# Benchmark YOLOv8n speed and accuracy on the COCO8 dataset for all all export formats
yolo benchmark model=yolov8n.pt data=coco8.yaml imgsz=640

Best Practices when using NVIDIA Jetson

When using NVIDIA Jetson, there are a couple of best practices to follow in order to enable maximum performance on the NVIDIA Jetson running YOLOv8.

1. Enable MAX Power Mode

   Enabling MAX Power Mode on the Jetson will make sure all CPU, GPU cores are turned on.

   sudo nvpmodel -m 0
   

2. Enable Jetson Clocks

   Enabling Jetson Clocks will make sure all CPU, GPU cores are clocked at their maximum frequency.

   sudo jetson_clocks
   

3. Install Jetson Stats Application

   We can use jetson stats application to monitor the temperatures of the system components and check other system details such as view CPU, GPU, RAM utilization, change power modes, set to max clocks, check JetPack information

   sudo apt update
   sudo pip install jetson-stats
   sudo reboot
   jtop
   

Next Steps

Congratulations on successfully setting up YOLOv8 on your NVIDIA Jetson! For further learning and support, visit more guide at Ultralytics YOLOv8 Docs!

Comments