A comprehensive deep learning project implementing and comparing two CNN architectures for CIFAR-10 image classification, featuring advanced training techniques and thorough performance analysis.

This project implements two distinct CNN architectures to classify images from the CIFAR-10 dataset:

• Simple CNN: A straightforward 3-layer convolutional network
• Improved CNN: An advanced architecture with residual blocks, progressive dropout, and regularization techniques

The project includes comprehensive evaluation metrics, overfitting analysis, and visualization tools to compare model performance.

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

• ✈️ Airplane, 🚗 Car, 🐦 Bird, 🐱 Cat, 🦌 Deer

• 🐕 Dog, 🐸 Frog, 🐴 Horse, 🚢 Ship, 🚛 Truck

• Training set: 50,000 images

• Test set: 10,000 images

• 3 convolutional layers (32 → 64 → 128 channels)
• Batch normalization and dropout for regularization
• MaxPooling for dimensionality reduction
• Fully connected layers for classification

• Residual blocks for better gradient flow
• Progressive dropout strategy (0.05 → 0.5)
• Global Average Pooling to reduce parameters
• Enhanced regularization to combat overfitting
• Adaptive learning with cosine annealing

• Combined Loss Function: Focal Loss + Label Smoothing Cross Entropy
• Data Augmentation: Random flips, rotations, color jittering
• Progressive Dropout: Dynamic dropout rates during training
• Cosine Annealing: Learning rate scheduling

• Training/validation loss and accuracy tracking
• Confusion matrix analysis
• Per-class performance metrics
• Overfitting detection and analysis

Latest comprehensive evaluation showing training progress, validation metrics, and final test accuracy for both models.

Analysis of training vs validation performance to detect and measure overfitting in both architectures.

pip install -r requirements.txt

• PyTorch ≥ 2.0.0
• torchvision ≥ 0.15.0
• matplotlib ≥ 3.5.0
• seaborn ≥ 0.11.0
• scikit-learn ≥ 1.0.0
• tqdm ≥ 4.60.0

Train and evaluate models:

python cifar10_main.py

CIFAR-10/
├── cifar10_main.py           # Main training and evaluation script
├── cifar10_colab.py          # Google Colab optimized version
├── CNN_simple.py             # Simple CNN architecture
├── CNN_improved.py           # Advanced CNN with residual blocks
├── colab_git_training.py     # Colab git integration
├── requirements.txt          # Project dependencies
├── data/                     # CIFAR-10 dataset storage
├── *.png                     # Generated evaluation plots
└── README.md                 # This file

The improved CNN architecture demonstrates:

• Better generalization through residual connections
• Reduced overfitting via progressive dropout
• Improved convergence with advanced loss functions
• Enhanced performance on challenging classes (cats, birds, dogs)

• Smart Loss Function: Combines Focal Loss for hard examples with Label Smoothing for better generalization
• Adaptive Training: Dynamic dropout rates and learning rate scheduling
• Comprehensive Analysis: Detailed overfitting detection and per-class performance metrics
• Production Ready: Clean, modular code with proper documentation

This project is open source and available under the MIT License.