Shape Classifier Demo

Context

I wanted a model that can tell what shape someone drew.

Approach

• Used Google's QuickDraw sketches and built train/validation splits.
• Trained a small ResNet18 classifier in PyTorch.
• Deployed a CPU-only AWS Lambda endpoint so the browser can request predictions.

Impact

• About 90% accuracy on five shapes: circle, triangle, square, hexagon, and octagon.
• After warm-up (~10 seconds), predictions return in under a second.

Data and Training

I trained a small classifier on Google’s QuickDraw sketches to recognize five basic shapes from a simple black-and-white drawing.

• Built class-balanced train/validation splits from QuickDraw categories.
• Used ResNet18 to keep the model small and fast.
• Exported plain weights (`model.pt`) so inference stays simple.

Inference API

• Implemented an AWS Lambda handler that takes a base64 image and returns a class plus confidence.
• Kept the API browser-friendly (CORS and small JSON payloads).
• Ran CPU-only inference so it works well in serverless environments.

Demo UX

• Canvas drawing UI with a clear submit step and a confidence bar.
• Clear status messaging during warm-up and cold starts.
• Preprocessing tuned so messy real-world strokes still work.

What I'd Improve

• Calibrate confidence (for example, temperature scaling) so scores are more reliable.
• Expand to more shapes and add augmentation for stroke width and incomplete outlines.
• Create a small human-drawn validation set to measure real-world performance.

Links

GitHub Live Demo