doodle-digit-classifier

Train a model to recognize hand-written digits using the MNIST dataset and Keras and TensorFlow, convert to a TensorFlowJS model and import it to a webpage to classify digits drawn by the user!

Deployment

To run this app locally you could set up a node server to host the tfjs model files (as the TensorFlow loadLayersModel function expects to retrieve a model over http request). I opted to use the npm module http-server for simplicity.

Steps using npm:

1. Clone the directory and run npm install
2. Run npm start.
3. Open the index.html file using your favourite browser and have fun!

Steps to set up manually:

1. Clone the directory and install http-server using: npm install http-server -g
2. Navigate to the project directory on your command line/terminal and deploy the server using: http-server -c1 --cors . or http-server -c-1 -p 8080
3. Open the index.html file using your favourite browser and voila!

Model Development

This model script is developed for Python 3.6.8. To run the build_model.py file it is recommended to either use a Docker container or a package manager such as Conda to ensure the correct Python version and to install dependencies.

Docker

The following Docker image will set up a container with the correct Python version and all required dependencies:

docker pull evenchange4/docker-tfjs-converter

Running it with the following command will set the output format for the TensorFlow models:

docker run -it --rm \ -v "$PWD/python:/python" \ evenchange4/docker-tfjs-converter \ tensorflowjs_converter --input_format=keras python/output/model.h5 python/output/model-tfjs

Conda

To train the model using Conda first ensure the Python version is correct: conda install python=3.6.8

then install the recquired packages, e.g:

pip install --user tensorflowjs