Distil ResNet50 to MobileNet on Raspberry Pi 5: Compare Model Size, Speed, and Accuracy

Every part needed, pre-tested for compatibility, with an AI build companion trained on this exact project. Shipped from Bengaluru in 3-5 days.

Run a full knowledge distillation pipeline on a Raspberry Pi 5 — train a MobileNet student to mimic a ResNet50 teacher, then benchmark both models on the edge. You’ll produce a detailed report comparing model size (MB), inference latency (ms), and classification accuracy, gaining hands-on insight into the real-world tradeoffs of AI model compression for embedded systems. This kit turns a research concept into a repeatable experiment right on your desk.

What You'll Build

A complete knowledge distillation experiment: you’ll set up a PyTorch training environment on Pi 5, implement a distillation loss combining hard and soft targets, train the student model, and evaluate both teacher and student on a standard image classification dataset (CIFAR-10/100). The NVMe SSD provides fast storage for dataset caching and model checkpoints, ensuring the Pi’s limited RAM isn’t a bottleneck. By the end, you’ll have a side-by-side comparison of a large convolutional network and its compact distilled version, along with a clear understanding of when the tradeoff makes engineering sense.

What You'll Learn

• Implement knowledge distillation from scratch in PyTorch on an ARM64 device, including temperature-scaled soft targets and loss weighting
• Benchmark inference speed on Pi 5’s CPU and built-in VideoCore VI GPU using torchprof or custom timing loops
• Quantify model memory footprint – RAM usage during inference and storage on the SSD – for both teacher and student
• Analyse accuracy loss versus compression ratio to inform deployment decisions for real-world edge AI systems

Kit Contents

Why Buy This Kit Instead of Sourcing Parts Separately

Who This Kit Is For

This advanced kit is designed for Indian engineering students and researchers tackling AI model optimisation for edge devices. B.Tech final-year students in CSE, ECE, or AI/ML will find it ideal for capstone projects, while M.Tech scholars can use it to prototype and benchmark distillation algorithms. Smart India Hackathon participants working on “AI for resource-constrained devices” problem statements will gain a ready-to-run experimental setup. Even industry professionals exploring TinyML will appreciate the Pi 5’s ability to run real training and inference locally.

Built and Backed by Compoden

Every Compoden kit ships with an AI build companion trained on this exact project — accessible via a QR code on the box, with WhatsApp and email backup. We've spent 10 years building projects for makers, schools, and institutions across India. If a part fails because of a manufacturing defect, replace it free within 7 days.