CIFAR-10 CNN Classifier with PyTorch

A custom convolutional neural network (MyNet) built from scratch using PyTorch to classify images from the CIFAR-10 dataset. This project includes a complete training pipeline, data augmentation, model checkpointing, visualization of results, and a confusion matrix for detailed performance analysis. The model achieved 87% accuracy on the test dataset.

---

📌 Features

• Custom CNN (MyNet) with 6 convolutional layers, BatchNorm, and Dropout
• Stratified train/validation split using StratifiedShuffleSplit
• Advanced data augmentation: AutoAugment, RandomCrop, and ColorJitter
• Training loop with real-time loss and accuracy tracking
• Learning rate scheduling with StepLR
• Final evaluation on the test dataset
• Visualizations: training vs. validation loss, accuracy, and confusion matrix

---

🏗️ Model Architecture

• Convolutional Layers: 6 layers with increasing depth (8 → 256 channels)
• Pooling: MaxPooling applied after specific layers
• Batch Normalization: Applied to the first 5 convolutional layers
• Activation Function: ReLU used throughout the network
• Classifier: Global Average Pooling → Fully Connected (512) → Fully Connected (10)
• Dropout: 50% dropout before the final fully connected layer

---

🧪 Data Pipeline

• Dataset: CIFAR-10, automatically downloaded using torchvision
• Splits:
  • Training: 80%
  • Validation: 20% (stratified)
  • Test: Official CIFAR-10 test set
• Augmentation:
  • AutoAugment policy for CIFAR-10
  • Random cropping and flipping
  • Color jitter for brightness, contrast, and saturation

---

🚀 How to Run

1. Install dependencies:

   pip install torch torchvision matplotlib numpy scikit-learn tqdm

2. Train the model:

   python3 MyNet.py

---

📊 Visual Output

🔻 Loss Plot

🔺 Accuracy Plot

📉 Confusion Matrix

---

📁 Project Structure

├── MyNet.py                     # Main script containing model, training, etc.
├── data/                         # CIFAR-10 dataset (auto-downloaded)
├── figures/                      # Directory for saved plots
│   ├── loss_fig.png              # Loss plot
│   ├── accuracy_fig.png          # Accuracy plot
│   └── confusion_matrix.png      # Confusion matrix
├── cifar_mynet_final.pt          # Final trained model
├── cifar_mynet_epoch_*.pt        # Model checkpoints
└── README.md                     # Project documentation