Skip to content

Projects 5: Neural Networks

CS50's Introduction to Artificial Intelligence with Python

Project — Traffic

Build a CNN to classify road signs using the German Traffic Sign Recognition Benchmark (GTSRB).

Background

Autonomous vehicles need to identify road signs from camera images. The GTSRB dataset contains thousands of images of 43 different road sign categories.

This project uses TensorFlow to design, train, and evaluate a convolutional neural network for this classification task.

Getting Started

pip3 install -r requirements.txt  # opencv-python, scikit-learn, tensorflow

Dataset structure: 

gtsrb/
  0/   ← 43 subdirectories, one per sign category
  1/
  ...
  42/

Place the gtsrb folder inside the project directory.

Code Structure

traffic.py — main() orchestrates the workflow (pre-built): 1. Load data via load_data 2. Split into training and test sets 3. Get compiled model via get_model 4. Train, evaluate, optionally save

Specification

Implement 2 functions:

load_data(data_dir)

Returns (images, labels) tuple: - Read images using OpenCV (cv2) - Resize all images to IMG_WIDTH × IMG_HEIGHT (constants defined in file) - Images as numpy arrays, labels as integers (0–42) - Use os.path.join for platform-independent paths

import cv2
img = cv2.imread(filepath)
img = cv2.resize(img, (IMG_WIDTH, IMG_HEIGHT))

get_model()

Returns a compiled TensorFlow/Keras model: - Input shape: (IMG_WIDTH, IMG_HEIGHT, 3) (RGB images) - Output layer: NUM_CATEGORIES units (43), softmax activation - Architecture is up to you — experiment with: - Number of conv/pooling layers - Filter counts and sizes - Hidden layer sizes - Dropout rates

model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(32, (3, 3), activation="relu",
                           input_shape=(IMG_WIDTH, IMG_HEIGHT, 3)),
    tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(128, activation="relu"),
    tf.keras.layers.Dropout(0.5),
    tf.keras.layers.Dense(NUM_CATEGORIES, activation="softmax")
])

model.compile(
    optimizer="adam",
    loss="categorical_crossentropy",
    metrics=["accuracy"]
)

Documentation Requirement

Create a README.md documenting your experiments: - What architectures did you try? - What worked well and what didn't? - Key observations

Constraints

  • Only modify load_data and get_model
  • Authorized imports: TensorFlow, OpenCV, scikit-learn, NumPy, Pandas
  • When submitting via Git, exclude the gtsrb directory (too large)

Running

python traffic.py gtsrb
python traffic.py gtsrb model.h5  # to save the model
check50 ai50/projects/2024/x/traffic
style50 traffic.py