🚀 Hybrid GPU Image Classification Pipeline

⚡ A high-performance deep learning pipeline combining PyTorch, FastAI, CuPy, and Numba — optimized for NVIDIA RTX GPUs — to preprocess and train on the CIFAR-10 dataset with maximum GPU acceleration.

---

🧠 Overview

This repository demonstrates a hybrid GPU pipeline that offloads image preprocessing and augmentation to the GPU using:

• 🧮 CuPy: Fast NumPy-like GPU array computations
• 🔬 Numba: Custom CUDA kernels for image brightness enhancement
• 🐍 FastAI: Rapid model training using PyTorch under the hood
• 🎮 RTX GPU Optimized: Designed to fully leverage your RTX GPU (20xx, 30xx, or 40xx series)

---

📊 Demo Output

🚀 Starting Hybrid GPU Image Classification Pipeline
Using device: cuda
Dataset loaded with batch size 128
Preprocessing sample batch with GPU kernels...
Starting training...
✅ FastAI training completed
⏱️ Time: 18.53 sec

---

🏗️ Pipeline Architecture

┌────────────────────┐
│ CIFAR-10 Dataset   │
└─────────┬──────────┘
          ↓
   FastAI DataLoaders
          ↓
  ┌────────────────────────────┐
  │ GPU Preprocessing Steps    │
  ├────────────────────────────┤
  │ 1. CuPy Normalization       │
  │ 2. Numba Brightness Kernel  │
  └────────────────────────────┘
          ↓
   FastAI ResNet-18 Model
          ↓
       Training Loop

---

⚙️ Installation

✅ Prerequisites

• NVIDIA RTX GPU with CUDA support
• Python 3.8+
• CUDA drivers installed and working
• CUDA-compatible versions of PyTorch and CuPy

📦 Install Dependencies

# For CUDA 11.8 and RTX GPUs
pip install cupy-cuda118

# Core packages
pip install torch torchvision fastai numba

💡 Choose the CuPy version that matches your installed CUDA toolkit:
See the CuPy install matrix.

---

🧪 How to Run

https://github.com/dragonpilee/Hybrid-GPU-Image-Classification-Pipeline.git
cd hybrid-gpu-classifier
python gpu_ml_pipeline.py

---

🔍 Code Highlights

🔧 gpu_normalize_images()

Normalizes RGB image tensors using CuPy directly on GPU memory.

💡 brightness_kernel (Numba)

Custom CUDA kernel that increases brightness pixel-wise on the GPU.

🧼 preprocess_batch()

Combines GPU normalization and augmentation, then converts back to PyTorch tensors.

🧠 train_model()

Initializes FastAI’s ResNet-18 model and performs training with one-cycle policy.

---

⚡ Performance Tips

Setting	Recommendation
GPU	NVIDIA RTX 2060/3060/4090 or higher
Batch Size	128–256 for optimal GPU utilization
Data Preprocessing	CuPy + Numba (already integrated)
Precision	Add mixed precision for even faster training
Memory Management	Use .half() and torch.cuda.amp for FP16

---

🚀 Future Improvements

• 🔁 Integrate GPU-accelerated preprocessing inside FastAI’s transform pipeline
• 🎨 Add more augmentation types (contrast, noise, rotation) via custom CUDA kernels
• 📈 Benchmark performance across different GPUs (RTX 2060 vs 4090)
• 💾 Add model saving, evaluation, and inference scripts

---

📜 License

This project is licensed under the MIT License.
Feel free to fork, modify, and use it in your own projects.

---

👨‍💻 Author

Alan Cyril Sunny
📧 alan_cyril@yahoo.com
🐙 GitHub

---

🌟 Show Your Support

If you found this useful, consider starring ⭐ the repo or sharing it with others!