This is a "You draw, AI guesses" MVP (Minimum Viable Product) that can recognize the sketch drawing on web canvas using a TensorFlow.js friendly CNN model.
This project utilizes Google's "The Quick Draw" dataset. The Quick Draw dataset is a collection of 50 million drawings across 345 categories, and if training the network model using the full dataset, the needed computation resource of GPU will be very huge for me. So as for demonstrating how to develop a AI enabled Web App by using TensorFlow full stack, I just sampled 10 categories and tens of thousands pictures per category from this dataset to train the CNN model, and achieved 94.87% accuracy after training 25 epochs on Google colab, then published the model as a web friendly model for TensorFlow.js based Web App.
To play the demo live https://ai.kyletiger.com/quickdraw-cnn, you need to use a modern web browser (e.g. Chrome, Safari, or Wechat on either web or mobile devices) that supports ES6 runtime. IE (or IE kernel based browsers) isn't supported!
- Server side:
TensorFlow + TensorLayer + Python - Client side:
TensorFlow.js + ES6 - Model UT:
TensorFlow.js + Node.js
Codes style of this MVP:
- More TF & TL (1.x) style: use more recent and decent TF APIs.
- More Pythonic: fully leverage the power of python.
- Readability (over efficiency): Since it's for instruction purposes, we prefer readability over others.
- Understandability (over everything): Understanding DL key concepts is the main goal of this code.
Copyright (C) 2000-2018 Lisong Zheng, 郑立松
The binaries and source code of this project can be used according to the Apache License, Version 2.0.