Lab01: Handwritten Digit Recognition with AI

Example Project Path: ai_labs/Lab01_HandwrittenDigits

Introduction

In this lab, you will train a deep learning model to recognize handwritten digits using Jupyter Notebook (inside Docker) on Windows 11, perform real-time digit recognition on Windows, and deploy the trained model to a Raspberry Pi 5 with a Hailo8 AI HAT and camera.

Expected Outcomes

• Understand the workflow for training and deploying an AI model.
• Use Docker and Conda for Python development environments.
• Run live digit recognition from webcam/camera on Windows and Raspberry Pi 5.
• Experience hardware-accelerated AI inference with the Hailo8 module.

Hardware Requirements

• Windows 11 PC (Intel/AMD)
• Raspberry Pi 5 (with Hailo8 AI HAT)
• Raspberry Pi camera module
• Optional: SSD for Pi 5

Part I: Train the AI Model on Windows 11

Step 1: Download and Prepare the Lab Files

1. Download and Extract Lab Files
   • Download the provided lab zip file and extract it to:
     C:\ai_labs\Lab01_HandwrittenDigits
2. Ensure your folder structure is:

   
   Lab01_HandwrittenDigits/
   |-- docker/
   │   |-- Dockerfile
   │   `-- docker-compose.yml
   |-- notebooks/
   │   `-- train_model.ipynb
   |-- src/
   │   |-- camera_single_digit.py
   │   `-- camera_multiple_digits.py
   |-- saved_model/
   │   `-- (model files will be saved here)
   |-- data/
   |   `-- (optional: for data storage)
   `-- requirements.txt
   
   

Step 2. Create and Launch a Docker Container for Jupyter

1. Open a terminal and change to the docker folder:
   

   cd C:\ai_labs\Lab01_HandwrittenDigits\docker

2. Build and run the Docker container:
   

   docker compose up --build

3. Open Jupyter Lab in your browser at:
   http://127.0.0.1:8888/lab

Step 3. Train and Export the Model

1. In Jupyter Lab, open notebooks/train_model.ipynb.
2. Follow the instructions to complete any missing code as directed by your instructor.
3. Set epochs=16 in the training code. (More epochs improve model accuracy on small datasets.)
4. Click Run All to execute all cells.
5. After completion, check that handwritten_digit_model.h5 appears in the saved_model folder.

Part II: Convert Model and Test on Windows 11

Step 1. Prepare Conda Environment for Model Conversion

1. Open Anaconda PowerShell Prompt.
2. Change to the project directory:
   

   cd C:\ai_labs\Lab01_HandwrittenDigits

3. Create and activate the environment:
   

   conda create --prefix .\.conda\envs\tf310 python=3.10
   conda activate .\.conda\envs\tf310

4. Install required packages:
   

   conda install -c conda-forge tensorflow numpy=1.26 opencv

Step 2. Convert the Keras Model (.h5) to TensorFlow Lite (.tflite)

1. Change to the src folder:
   

   cd src

2. Copy the trained model:
   

   copy ..\saved_model\handwritten_digit_model.h5 .

3. Create a script named convert_to_tflite.py with the following content:

   
   # convert_h5_tflite.py
   import tensorflow as tf
   model = tf.keras.models.load_model('../saved_model/handwritten_digit_model.h5')
   converter = tf.lite.TFLiteConverter.from_keras_model(model)
   tflite_model = converter.convert()
   with open('../saved_model/handwritten_digit_model.tflite', 'wb') as f:
       f.write(tflite_model)
   print("Conversion complete: handwritten_digit_model.tflite")
   

4. Run the conversion:

   python convert_to_tflite.py

5. Confirm that handwritten_digit_model.tflite is created.

Step 3. Test Model Inference on Windows

1. Run single-digit recognition:

   python camera_single_digit.py

   Hold a single digit up to your webcam. Press q to quit.
2. Run multiple-digit recognition:

   python camera_multiple_digits.py

   Show multiple digits. Press q to quit.
3. Deactivate Conda:

   conda deactivate

Part III: Deploy and Run on Raspberry Pi 5

Prerequisites

• Docker and Conda installed on Pi 5
• Hailo8 AI HAT and Pi camera attached
• Run sudo apt install hailo-all
• Hailo Python SDK/RT installed under /usr/lib/python3/dist-packages
• Pi camera tested and functional

Step 1. Create Project Directory

• In your terminal:

  
  mkdir -p /ssd/ai_labs/Lab01_HandwrittenDigits
  cd /ssd/ai_labs/Lab01_HandwrittenDigits
  

Step 2. Transfer the Trained Model

• Copy handwritten_digit_model.tflite from Windows to Pi 5, for example (adjust [PI_IP]):

scp C:\ai_labs\Lab01_HandwrittenDigits\src\handwritten_digit_model.tflite pi@[PI_IP]:/ssd/ai_labs/Lab01_HandwrittenDigits/

Step 3. Install All Necessary Packages for System Python

1. Update and install system Python packages:
   sudo apt update
sudo apt install python3-picamera2 python3-opencv python3-numpy
2. Install TensorFlow Lite Runtime (ignore warning about --break-system-packages):
   sudo pip3 install tflite-runtime --break-system-packages
3. Verify the installation:

   /usr/bin/python3 -c 'import cv2, numpy; from picamera2 import Picamera2; from tflite_runtime.interpreter import Interpreter; print("All packages OK!")'
       

   You should see: All packages OK!

Step 4. (Optional) Check Hailo Packages/Drivers

• List Hailo packages (if using Hailo for acceleration):

ls /usr/lib/python3/dist-packages | grep hailo
ls /usr/lib/python3/dist-packages/hailo_platform/drivers

Step 5. Execute the Python Code for Camera Digit Recognition

1. Use the Picamera2-based script, for example, pi_camera_single_digit_tflite_picamera2.py.
2. Run the script with system Python: /usr/bin/python3 pi_camera_single_digit_tflite_picamera2.py
3. Hold a handwritten digit in front of the camera. The prediction will show in a window. Press q to exit.

Appendix: Useful Commands

• Check installed Python version: python3 --version
• Check installed packages: pip3 list
• Update Raspberry Pi OS: sudo apt update && sudo apt upgrade
• Check camera connection: libcamera-hello (or rpicam-hello on Pi 5 with rpicam-apps)

Appendix A: Useful Docker Commands

Appendix B: Useful Conda Commands

Notes:

• Docker and Conda are assumed pre-installed.
• Jupyter notebook: train_model.ipynb in notebooks/
• Recognition scripts: camera_single_digit.py and camera_multiple_digits.py in src/
• On Pi, use conda for main packages, pip for TensorFlow Lite (tflite-runtime).