Overview
What this challenge is about.
You receive a trained FP32 PyTorch person-detector (mAP 0.74 on a 5k validation set) plus a calibration dataset of 500 unlabeled doorbell frames. Convert to ONNX, then apply post-training INT8 quantization with both per-tensor and per-channel schemes. Benchmark on a Raspberry Pi 3B+ (Cortex-A53 proxy for the production SoC) using ONNX Runtime. Report mAP, mean inference latency, peak RAM, and model file size for FP32, INT8 per-tensor, and INT8 per-channel. Write a one-page memo: ship as-is, or invest 2 sprints in QAT.
The Brief
What you'll do, and what you'll demonstrate.
Quantize a person-detection model to INT8 for a Cortex-A53 target without dropping accuracy below the product threshold, and decide whether QAT is needed.
Earning criteria — what you'll demonstrate
- Apply post-training quantization to a real vision model
- Compare per-tensor vs. per-channel quantization schemes
- Benchmark inference on a real ARM target (or documented proxy)
- Reason about accuracy/latency/memory trade-offs for shipping
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
Machine Learning Engineer
Shipping a quantized vision model with honest on-device benchmarks is exactly the day-one work of an MLE at a consumer-IoT or edge-AI company.
This challenge sharpens
- quantization
- model-optimization
- edge-inference
MLOps Engineer
Reproducible export + calibration + benchmark scripts mirror the MLOps craft of building reliable model-shipping pipelines.
This challenge sharpens
- onnx
- benchmarking
- edge-inference
AI Engineer
Translating a model and a hardware constraint into a ship/no-ship recommendation is core AI-engineer work at any product-led AI startup.
This challenge sharpens
- quantization
- pytorch
- benchmarking