🚀 Installing PyTorch and TorchVision on JetPack 6.1

For NVIDIA Jetson Orin AGX Developer Kit (64GB)

---

The NVIDIA Jetson AGX Orin is a cutting-edge platform designed for advanced AI and computer vision applications. It delivers up to 275 TOPS (Tera Operations Per Second), powered by the Ampere GPU architecture with 2048 CUDA cores and 64 Tensor Cores, enabling real-time processing of complex vision models like YOLO and Mask R-CNN.

---

---

🔍 Overview

Setting up PyTorch and TorchVision on JetPack 6.1 can be a challenging task due to limited and often outdated documentation. This guide will walk you through installing PyTorch and building TorchVision from source to get your ## NVIDIA Jetson Orin AGX Developer Kit (64GB) ready for AI/ML tasks.

---

🤔 Why This Guide?

I faced several challenges while setting up my device for YOLO-based real-time object detection. The lack of accurate guidelines from official sources, even on NVIDIA forums, made it a time-consuming process. Here's a streamlined guide based on my experience to save you the hassle.

---

🛠 Steps to Install PyTorch and TorchVision on Ubuntu 22.04 Jetson Orin AGX Developer Kit

1️⃣ Prerequisites

Update your system and install required dependencies:

sudo apt-get update && sudo apt-get upgrade -y
sudo apt-get install -y python3-pip libjpeg-dev libpng-dev libtiff-dev

Set up a Python virtual environment (recommended):

python3 -m venv ~/env
source ~/env/bin/activate

2️⃣ Install PyTorch

Download the compatible PyTorch wheel from NVIDIA: https://developer.nvidia.com/embedded/downloads

Visit NVIDIA PyTorch Downloads. Select the PyTorch wheel for JetPack 6.1: Example file: **torch-2.5.0a0+872d972e41.nv24.08-cp310-cp310-linux_aarch64.whl ** Install PyTorch:

pip install /path/to/torch-2.5.0a0+872d972e41.nv24.08-cp310-cp310-linux_aarch64.whl

3️⃣ Build TorchVision from Source

Since NVIDIA doesn’t provide a prebuilt TorchVision wheel for JetPack 6.1, you’ll need to build it from source. Clone the Repository:

git clone https://github.com/pytorch/vision.git
cd vision

Checkout the Compatible Version:

Make sure the TorchVision version aligns with your PyTorch version:

git checkout tags/v0.20.0

Build and Install TorchVision:

python3 setup.py install

⚠️ Note: Ensure you’ve installed dependencies (libjpeg-dev, libpng-dev, libtiff-dev) before building.

4️⃣ Verify Installation

Run this script to confirm the setup:

import torch
import torchvision

print("Torch Version:", torch.__version__)
print("TorchVision Version:", torchvision.__version__)
print("CUDA Available:", torch.cuda.is_available())

if torch.cuda.is_available():
print("CUDA Device:", torch.cuda.get_device_name(0))

✔️ Expected Output:

Torch Version: 2.5.0a0+872d972e41.nv24.08
TorchVision Version: 0.20.0
CUDA Available: True
CUDA Device: Orin

⚙️ Why Build TorchVision from Source?

Building TorchVision from source ensures:

Compatibility with your specific PyTorch version. Flexibility to customize or optimize for your needs.

🐛 Troubleshooting

Common Issues:

Version Mismatch: Verify TorchVision matches your installed PyTorch version. Use git checkout to get the correct version from GitHub.

Missing Dependencies: Ensure dependencies like libjpeg-dev, libpng-dev, and libtiff-dev are installed:

sudo apt-get install libjpeg-dev libpng-dev libtiff-dev

**Build Errors: ** Update setuptools and pip:

pip install --upgrade pip setuptools

+ Extra Link to Download Latest Packages:

http://jetson.webredirect.org/jp6/cu126

🎉 Conclusion

Follow this guide to efficiently set up PyTorch and TorchVision on JetPack 6.1, leveraging the full potential of NVIDIA Jetson Orin AGX for AI/ML projects. Save and share this guide with others to streamline their setup process. Happy coding! 🚀

---

I successfully set up the NVIDIA Jetson AGX Orin with YOLO for real-time object detection and safety monitoring. Leveraging the powerful Jetson AGX Orin's 275 TOPS performance, the system processes live video streams to detect essential safety features like helmets, vests, and goggles with impressive accuracy and low latency.

Highlights of the Setup:

Real-Time Detection:

Objects such as "Helmet: No," "Vest: No," and "Goggles: Yes" are identified instantly with bounding boxes and confidence scores, ensuring quick decision-making.

YOLO Integration:

Custom-trained YOLO models were deployed seamlessly using the NVIDIA JetPack SDK, optimized with TensorRT for edge AI applications.

Efficiency and Scalability:

The setup supports multiple video streams, making it suitable for industrial environments, ensuring worker compliance with PPE (Personal Protective Equipment) regulations.

Custom Dataset:

A personalized dataset was created and trained on the Jetson, enhancing the detection accuracy for real-world scenarios.

This achievement demonstrates the power of NVIDIA Jetson AGX Orin in combining AI and computer vision for critical real-time applications like workplace safety monitoring, reducing risks, and improving operational efficiency. 🚀

---

💡 Author

👨‍💻 Azimjon Akhtamov AI Research Assistant | Graduate Computer Science Student at CBNU | South Korea

☘️LinkTree: https://linktr.ee/azimjaantech

⭐ Contribute

Fork this repository and submit pull requests for improvements or additional issues. Let’s help the community together! 💪