What's your number?

Toy project implementing the "Hello Word" of Machine Learning, namely the MNIST Machine Learning Challenge. The project aims to experiment with the latest APIs (e.g. TensorFlow v2.0).

You can see the demonstration of the UI. As you can see, the model makes mistakes, as it is not a perfectly trained one.

How to launch the drawing UI

$ docker build -t ui_image .
$ docker run --name ui_container \
             --rm \
             -it \
             -p 3000:3000 \
             ui_image

Once the JavaScript server is up within ui_container, the user reach the UI under http://localhost:3000.

How to train a new model

$ docker build -t train_image train_context
$ docker run --name train_container \
             --rm \ 
             -it \
             -p 8888:8888 \
             --mount src=$(pwd)/model,target=/home/docker/model,type=bind \
             train_image

docker run starts a Jupyter server on :8888 (hence the port publishing argument of docker run). The server can be reached at the command line provided URL (URL is provided as a result of docker run).

Above configuration includes plotly extension installation into jupyterlab. This installation step takes quite long to execute, besides even fails if not enough resources are assigned to the Docker Engine. plotly is a nice-to-have visualisation tool to draw pretty plots – therefore if a quicker-to-build configuration is required, plotly is recommended to be replaced with matplotlib. Within this project, matplotlib would be more than satisfactory, default configuration includes plotly only for the nicer plots.

After docker run gets executed, the home folder of train_container will look like this:

home
└── docker
    ├── model
    │   ├── group1-shard1of1.bin
    │   ├── model.h5
    │   └── model.json
    └── train_model.ipynb

• model is the folder of
  • the .h5 format of the model (format before conversion),
  • the .bin+.json format of the model (format after conversion).
• train_model.ipynb is the Jupyter Notebook with which the user can train a new model.

To train a new model, open the Jupyter server within your browser, open train_model.ipynb and execute its cells.

How to convert the newly trained model

$ docker build -t convert_image convert_context
$ docker run --name convert_container \
             --rm \ 
             -it \
             --mount src=$(pwd)/model,target=/home/docker/model,type=bind \
             convert_image

# within convert_container
$ tensorflowjs_converter --input_format=keras model/model.h5 model

We override the default model with the above conversion. If ui_container is still running, restarting its JavaScript server will start using the newly converted model.