🧠 DeepVision Suite

Team Members:

• 1. Nuthalapati sai Bharath Kumar
• 2. Pranjal Prakash Pandey
• 3. Shriprasad Patil

A comprehensive deep learning image classification system built with PyTorch, featuring model interpretability tools and MLOps practices.

🚀 Overview

DeepVision Suite is a production-ready image classification system trained on the Tiny-ImageNet dataset (200 classes). It goes beyond just achieving high accuracy—it focuses on Model Interpretability and includes tools for understanding why the model makes certain decisions.

Key Features

• 🏗️ Custom ResNet Architecture with residual connections
• 📊 Model Visualization Suite (Grad-CAM, Feature Maps, Filter Visualization)
• 🔄 MLOps Integration (W&B tracking, GitHub Actions CI/CD)
• 🚀 Production Ready (FastAPI serving, ONNX export)

📁 Project Structure

DeepVisionSuite/
├── train.py                 # Main training script
├── visualize_model.py       # Visualization utilities (Grad-CAM, etc.)
├── requirements.txt         # Python dependencies
├── configs/
│   └── train_config.yaml    # Training configuration
├── scripts/
│   └── export_onnx.py       # ONNX model export
├── serve/
│   ├── app.py               # FastAPI inference server
│   └── Dockerfile           # Container for deployment
├── tests/
│   └── test_model.py        # Unit tests
├── visualizations/          # Generated visualization outputs
└── .github/workflows/       # CI/CD pipelines

🛠️ Installation

# Clone the repository
git clone https://github.com/Shriprasad-P/DeepVisionSuit.git
cd DeepVisionSuit

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

🎯 Training

Basic Training

python train.py --epochs 10

With Weights & Biases Tracking

python train.py --epochs 20 --wandb

Configuration Options

Argument	Default	Description
--epochs	10	Number of training epochs
--lr	0.001	Learning rate
--wandb	False	Enable W&B experiment tracking

📊 Model Visualization

Generate interpretability visualizations:

python visualize_model.py

This creates:

• Architecture Diagram — Network structure visualization
• Conv1 Filters — Learned first-layer features
• Feature Maps — Layer-by-layer activations
• Grad-CAM — Class activation heatmaps

🚀 Deployment

FastAPI Server

cd serve
uvicorn app:app --reload --port 8000

Docker

cd serve
docker build -t deepvision-api .
docker run -p 8000:8000 deepvision-api

ONNX Export

python scripts/export_onnx.py

🧪 Testing

pytest tests/ -v

🏗️ Architecture

The model uses a custom ResNet-style architecture with:

• Residual Connections — Prevents vanishing gradients
• Batch Normalization — Stabilizes training
• AdamW Optimizer — Weight decay regularization
• Cosine Annealing LR — Learning rate scheduling

📈 Results

Training on Tiny-ImageNet (200 classes, 64×64 images):

• Dataset: 100,000 training images / 10,000 validation images
• Architecture: Custom ResNet with 4 residual blocks

🔮 Future Improvements

• Vision Transformer (ViT) implementation
• TensorRT optimization for edge deployment
• Object detection with YOLO backbone
• Medical imaging fine-tuning

📄 License

This project is licensed under the MIT License.

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

---

Made with ❤️ using PyTorch