Improving CIFAR-10 Classification

A systematic approach to improving image classification accuracy on the CIFAR-10 dataset using PyTorch.

Overview

This project explores various techniques to improve deep learning model performance on CIFAR-10:

• Experiment tracking system for comparing different configurations
• GPU acceleration (optimized for RTX 5090)
• Systematic hyperparameter tuning
• Model architecture improvements
• Data augmentation strategies

Features

• Automated Experiment Tracking: Every training run is tracked with hyperparameters, accuracy, and timing
• GPU Support: Automatic device detection and optimization for CUDA-enabled GPUs
• Results Visualization: Color-coded comparison tables showing all experiments
• Persistent Storage: CSV export/import for tracking experiments across sessions

Getting Started

Prerequisites

pip install torch torchvision numpy matplotlib pandas

Running Experiments

1. Open image_classifier.ipynb in Jupyter
2. Run cells 1-3 to setup environment and load data
3. Modify hyperparameters in cell 7 for different experiments
4. Run training cells (7-8)
5. Evaluate and track results (cells 13-15)

Experiment Tracking

Each experiment automatically records:

• Model architecture
• Hyperparameters (learning rate, batch size, optimizer, etc.)
• Training time
• Test accuracy
• Device used (CPU/GPU)

View all experiments in a comparison table with color-coded accuracy metrics.

Dataset

CIFAR-10: 60,000 32x32 color images in 10 classes

• 50,000 training images
• 10,000 test images
• Classes: plane, car, bird, cat, deer, dog, frog, horse, ship, truck

Project Structure

├── image_classifier.ipynb    # Main notebook with experiment tracking
├── data/                      # CIFAR-10 dataset (auto-downloaded)
├── experiment_results.csv     # Saved experiment history
└── cifar_net.pth             # Saved model weights

Improvement Strategies

Quick Wins

• Increase epochs (50-100+)
• Add data augmentation
• Increase batch size
• Use Adam optimizer

Advanced Techniques

• Deeper architectures (ResNet)
• Batch normalization
• Learning rate scheduling
• Ensemble methods

Results

Track your progress! Each run is saved to experiment_results.csv and can be visualized in the results table.

License

MIT License

Acknowledgments

• CIFAR-10 dataset from Canadian Institute for Advanced Research
• PyTorch team for the excellent framework