This repo uses the tensorflow provided dataset MNIST.
The model is trained using Python and utilized in Javascript using the library tensorflowjs.
Two deep learning models were trained, "Convolutional Neural Network" and a "Regular Fully-connected Neural Network".
The CNN was primarily used for better accuracy.
The application is able to predict which digit number (from 0 to 9) the user has drawn on the canvas. This is achieved by machine learning algorithms and Javascript implementations.
More, user is able to make corrections to the prediction, providing a correct answer. The model would then adjust itself.
This must be ran in a server, i.e. a LocalHost. You should open code/main.html in browser through your IDE.
Although the model is auto-converted in the python script, you may want to convert it yourself manually sometime. Under Digit-Recognition folder, run the following bash script in your terminal
tensorflowjs_converter --input_format keras \
model/cnn/cnn.model \
model/cnn/json
Remember to rename the manually converted model model.json to cnn.json.