CIFAR-10 Image Classification (CNN Project)

This repository contains a Convolutional Neural Network (CNN) model built using TensorFlow/Keras to classify images from the CIFAR-10 dataset into 10 categories.

Features

• Loads and preprocesses CIFAR-10 dataset
• Custom train/validation/test split
• CNN with 3 convolutional blocks
• Model training with accuracy/loss tracking
• Evaluation using test accuracy and classification report
• Includes experiments with:
  • More training epochs
  • Smaller dense layer size
  • Deeper CNN architecture

Project Structure

machine.py      # Main script for model creation, training, evaluation
Report.pdf      # Full project report with experiment results and analysis
README.md       # Project documentation

Model Architecture (Baseline)

• Conv2D(32) → MaxPooling2D
• Conv2D(64) → MaxPooling2D
• Conv2D(128) → MaxPooling2D
• Flatten → Dense(128) → Dropout(0.5) → Dense(10 Softmax)

Requirements

Install the necessary dependencies:

pip install tensorflow numpy matplotlib scikit-learn

How to Run

Run the training and evaluation script:

python3 machine.py

This script will:

• Train the CNN model
• Display accuracy/loss curves
• Evaluate the model on the test dataset
• Print the classification report

Results (Summary)

• Baseline test accuracy: ~75%
• Additional experiments explored:
  • 25 training epochs
  • Dense layer reduced to 64 units
  • Added additional 256-filter convolutional block

(See Report.pdf for detailed graphs, metrics, and analysis.)

License

Created for CS 6840/4840 – Intro to Machine Learning.
For academic and educational use.