GAN

End-to-End CNN Model for Image Classification

A complete implementation of a Convolutional Neural Network (CNN) for image classification using PyTorch. This project includes data loading, model training, evaluation, and inference capabilities with comprehensive metrics tracking and visualization.

🌟 Features

• Modular Architecture: Clean separation of data, models, training, and utilities
• CIFAR-10 Dataset: Built-in support for CIFAR-10 with automatic downloading
• Data Augmentation: Configurable augmentation pipeline (random crops, flips, color jitter)
• Training Infrastructure:
  • Early stopping and learning rate scheduling
  • Model checkpointing (best and latest)
  • Comprehensive metrics tracking (accuracy, precision, recall, F1)
  • Training history visualization
• Evaluation Tools:
  • Confusion matrix generation
  • Per-class metrics analysis
  • Sample prediction visualization
• Inference: Easy-to-use script for making predictions on new images

📁 Project Structure

CNN1/
├── config.py                 # Configuration and hyperparameters
├── train.py                  # Main training script
├── evaluate.py               # Model evaluation script
├── inference.py              # Inference on new images
├── requirements.txt          # Python dependencies
├── data/                     # Data loading and preprocessing
│   ├── __init__.py
│   ├── dataset.py           # Dataset loaders
│   └── transforms.py        # Data transformations
├── models/                   # Model architectures
│   ├── __init__.py
│   ├── cnn.py              # CNN model definition
│   └── utils.py            # Model utilities
├── utils/                    # Training and evaluation utilities
│   ├── __init__.py
│   ├── trainer.py          # Trainer class
│   ├── metrics.py          # Metrics tracking
│   └── visualization.py    # Visualization functions
├── checkpoints/             # Saved model checkpoints
├── logs/                    # Training logs
└── results/                 # Evaluation results and plots

🚀 Installation

1. Clone the repository (or navigate to the project directory):

cd c:\CNN\CNN1

1. Install dependencies:

pip install -r requirements.txt

📊 Dataset

The project uses the CIFAR-10 dataset, which consists of:

• 60,000 32x32 color images in 10 classes
• 50,000 training images
• 10,000 test images
• Classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck

The dataset will be automatically downloaded on first run.

🎯 Usage

Training

Train the model with default settings:

python train.py

Train with custom parameters:

python train.py --epochs 50 --batch-size 64 --lr 0.0001

Training outputs:

• Model checkpoints saved to checkpoints/
• Training history saved to results/training_history.json
• Training plots saved to results/training_history.png

Evaluation

Evaluate the trained model on the test set:

python evaluate.py

Evaluate a specific checkpoint:

python evaluate.py --checkpoint checkpoints/best_model.pth

Evaluation outputs:

• Metrics saved to results/evaluation_metrics.json
• Confusion matrix saved to results/confusion_matrix.png
• Per-class metrics plot saved to results/per_class_metrics.png
• Sample predictions saved to results/sample_predictions.png

Inference

Make predictions on a new image:

python inference.py --image path/to/image.jpg --visualize

With custom parameters:

python inference.py --image path/to/image.jpg --checkpoint checkpoints/best_model.pth --top-k 3 --visualize --output results/prediction.png

⚙️ Configuration

All hyperparameters and settings can be modified in config.py:

Model Architecture

• conv_channels: [32, 64, 128, 256] - Channels for each conv block
• fc_hidden: 512 - Hidden units in FC layer
• dropout_rate: 0.5 - Dropout probability

Training Hyperparameters

• NUM_EPOCHS: 100
• BATCH_SIZE: 128
• LEARNING_RATE: 0.001
• OPTIMIZER: 'Adam' (options: Adam, SGD, AdamW)
• SCHEDULER: 'ReduceLROnPlateau'

Data Augmentation

• Random crop with padding
• Random horizontal flip
• Color jitter (brightness, contrast, saturation, hue)
• Optional random rotation

Early Stopping

• EARLY_STOPPING_PATIENCE: 15 epochs
• MIN_DELTA: 0.001

🏗️ Model Architecture

The CNN model consists of:

1. Convolutional Blocks (4 blocks):

   • Each block: Conv2D → BatchNorm → ReLU → Conv2D → BatchNorm → ReLU → MaxPool
   • Progressive channel increase: 3 → 32 → 64 → 128 → 256

2. Fully Connected Layers:

   • FC1: feature_size → 512 (with dropout)
   • FC2: 512 → 256 (with dropout)
   • FC3: 256 → 10 (output)

3. Regularization:

   • Batch normalization after each conv layer
   • Dropout (0.5) in FC layers
   • Weight decay (L2 regularization)

📈 Expected Results

With default configuration, you can expect:

• Test Accuracy: ~75-85% on CIFAR-10
• Training Time: ~2-3 hours on GPU, longer on CPU
• Model Size: ~5-10M parameters

🔧 Advanced Usage

Custom Dataset

To use a custom dataset, modify data/dataset.py:

def get_custom_loaders():
    # Implement your custom dataset loading
    pass

Model Customization

Modify the model architecture in models/cnn.py:

model = CNN(
    num_classes=10,
    conv_channels=[64, 128, 256, 512],  # Deeper network
    fc_hidden=1024,
    dropout_rate=0.3
)

Training Callbacks

The trainer supports:

• Early stopping
• Learning rate scheduling (ReduceLROnPlateau, StepLR, CosineAnnealing)
• Model checkpointing
• Metrics logging

📝 Metrics Tracked

During training and evaluation:

• Loss: Cross-entropy loss
• Accuracy: Overall classification accuracy
• Precision: Weighted average precision
• Recall: Weighted average recall
• F1 Score: Weighted average F1
• Per-class metrics: Individual metrics for each class
• Confusion Matrix: Detailed classification matrix

🐛 Troubleshooting

CUDA Out of Memory

Reduce batch size in config.py:

BATCH_SIZE = 64  # or 32

Slow Training

• Reduce NUM_WORKERS if CPU is bottleneck
• Use GPU if available
• Reduce model size (fewer channels/layers)

Poor Accuracy

• Increase training epochs
• Adjust learning rate
• Enable more data augmentation
• Try different optimizer/scheduler

📚 Dependencies

• Python 3.7+
• PyTorch 2.0+
• torchvision
• numpy
• matplotlib
• scikit-learn
• tqdm
• Pillow
• seaborn

🤝 Contributing

Feel free to:

• Report bugs
• Suggest features
• Submit pull requests
• Improve documentation

📄 License

This project is open source and available for educational purposes.

🙏 Acknowledgments

• CIFAR-10 dataset by Alex Krizhevsky
• PyTorch framework
• Community contributions and feedback

---

Happy Training! 🚀