Objective

The primary objective of this lab is to introduce students to the NVIDIA Jetson Orin Nano development platform and demonstrate how to build a real-time image classification system. Students will configure their devices in both graphical and headless modes, set up the development environment using Docker, and deploy a deep learning model for classifying hand gestures (thumbs up and thumbs down) using a live camera feed.

Learning Outcomes

Upon completion of this lab, students will be able to:

• Configure the Jetson Orin Nano board and connect via SSH in headless mode.
• Set up and launch JupyterLab using NVIDIA DLI Docker containers.
• Use iPython widgets within Jupyter notebooks for data collection via webcam.
• Fine-tune a pre-trained ResNet-18 model for binary image classification tasks.
• Evaluate and test model predictions in real time using a live webcam feed.
• Iterate and improve the model by collecting more diverse training data.
  

Lab Tasks

1. Hardware and OS Setup: Connect the Jetson Nano with monitor, mouse, and keyboard, or establish headless access using a Type-C data cable. Log into the system and connect to Wi-Fi.
2. Docker Environment Configuration: Create a persistent data directory and execute the docker_dli_run.sh script to pull and run the DLI AI container.
3. Launching Jupyter Notebook: Access the JupyterLab interface via 192.168.55.1:8888 and open the notebook classification_interactive.ipynb.
4. Data Collection: Use the notebook’s interactive widget to capture images for two classes — thumbs up and thumbs down — through a connected webcam.
5. Model Training: Train a modified ResNet-18 model with a final layer adjusted for binary classification. Monitor the training progress through live loss and accuracy metrics.
6. Live Testing: Evaluate the trained model by showing real-time hand gestures to the camera. Interpret prediction probabilities via slider widgets.
7. Model Improvement: Collect additional images with varying backgrounds, angles, and lighting conditions. Retrain themodel to improve accuracy and robustness.

Technologies Used

• Jetson Orin Nano (Ubuntu-based platform)
• Docker with NVIDIA runtime (nvcr.io/nvidia/dli/dli-nano-ai)
• JupyterLab with Python notebooks
• ResNet-18 (PyTorch-based)
• Webcam input and live video stream analysis

Expected Deliverables

• Screenshots of the Jupyter interface with trained model outputs.
• Brief written report including training accuracy and observations.
• Notes on data augmentation strategies used for improving model performance.