What Are You Drawing?
A ResNet50 model trained on Google's QuickDraw dataset guesses your sketches in real time - inference runs entirely in the browser with ONNX Runtime.
draw something
loading model…
What Are You Drawing? drops a trained ResNet50 directly into the page and asks it to recognise your sketches. Draw an object and watch the guess update live beneath the canvas, with a probability chart showing how confident the model is.
What it does
- Interactive drawing board rendered on a browser canvas
- Instantly guesses your drawing as you sketch
- Probability chart visualizes the model's prediction confidence
- One-click clear button to reset the canvas
How it works
Every stroke is debounced and the drawing is downscaled into the shape the model was trained on, then fed to the network - which runs on-device via ONNX Runtime, so nothing is uploaded. The model was trained in a Jupyter Notebook on Google's QuickDraw dataset to recognize basic shapes.
Tech highlights
- ONNX Runtime Web (WASM) for client-side inference - no server, no API key
- ResNet50 fine-tuned on Google's QuickDraw dataset
- Canvas 2D drawing with live prediction updates
Model training
The training pipeline lives in a Jupyter Notebook in the repo - it loads and preprocesses QuickDraw data, then trains the shape classifier from scratch.