LadyJ101/edge_benchmark

Benchmarking lightweight vision models for edge AI deployment under Nigeria-inspired constraints # **Edge Benchmark: Lightweight Vision Model Evaluation for Edge AI in Nigeria** This project benchmarks lightweight convolutional neural networks (CNNs) for image classification on CPU‑only, low‑cost edge devices, reflecting real deployment constraints common in Nigeria—limited power, low memory, and no GPU availability. The goal is to provide practical, reproducible insights into which models perform best on affordable hardware such as Raspberry Pi‑class boards, ARM devices, and low‑power embedded systems. --- ## 🎯 Research Goal Evaluate lightweight vision models under Nigeria‑inspired edge constraints, focusing on: - Latency (ms per image) - Accuracy (CIFAR‑10) - Model Size (MB) - Real‑world deployability on low‑resource hardware This benchmark answers: *Which lightweight model is most suitable for real‑world edge AI deployment in low‑resource environments?* --- ## 📦 Models Evaluated All models use PyTorch torchvision pretrained weights: - ShuffleNetV2 — optimized for speed - MobileNetV2 — balanced accuracy and efficiency - SqueezeNet — extremely small footprint - ResNet18 — heavier baseline for comparison --- ## 🧪 Benchmark Workflow All experiments were conducted in `notebooks/experiments.ipynb`. ### 1. Dataset - CIFAR‑10 (60,000 images, 10 classes) - Loaded via `torchvision.datasets.CIFAR10` ### 2. Metrics - Inference Latency (ms/image) - Accuracy (%) - Model Size (MB) ### 3. Hardware Constraints Simulated realistic Nigerian edge‑deployment conditions: - CPU‑only execution - ≤ 1 GB RAM - No GPU - Low‑power environment ### 4. Experiments Performed - Loaded pretrained models - Measured latency on CPU - Computed CIFAR‑10 accuracy - Calculated model size - Logged results for comparison --- ## 📁 Project Structure edge_benchmark/ │ ├── models/ # Model definitions ├── datasets/ # Dataset storage ├── scripts/ # Benchmarking + export scripts ├── notebooks/ # Experiments and testing └── results/ # Logs, outputs, and tables --- ## 📊 …