alexwickstrom / hera

Train/evaluate a Keras model, get metrics streamed to a dashboard in your browser.

Hera

Train/evaluate a Keras model, get metrics streamed to a dashboard in your browser.

Setting up

Step 1. Plant the spy

Install the package

    pip install heraspy

Add the callback

    herasCallback = HeraCallback(
        'model-key',
        'localhost',
        4000
    )

    model.fit(X_train, Y_train, callbacks=[herasCallback])

Step 2. Start the server

Git clone this repository, then run

    cd server
    npm install
    gulp # optional, for now the build file is kept track in git
    node build/server

Step 3. Start the dashboard

    cd client
    npm install
    npm start

Using RabbitMQ

By default hera uses socket.io for messaging - both from keras callback to server, and from server to dashboard. This is to minimize the number of things one needs to install before getting up and running with hera.

However, in production socket.io is outperformed by a number of alternatives, also it is good in general to decouple the server-client communication from the inter-process communitation (python -> node) so that each can be managed and optimized independently.

To demonstrate how this works Hera ships with the option to use rabbitMQ for interprocess communication. Here's how to use it.

In your model file

    from heraspy.callback import HeraCallback
    from heraspy.dispatchers.rabbitmq import get_rabbitmq_dispatcher

    herasCallback = HeraCallback(
        'model-key', 'localhost', 4000,
        dispatch=get_rabbitmq_dispatcher(
          queue='[my-queue]',
          amqps_url='amqps://[user]:[pass]@my-amqp-address'
        )
    )

In server/src/server.js

Replace the only line in the file with

    getServer({
        dispatcher: 'rabbitmq',
        dispatcherConfig: {
            amqpUrl: 'amqps://[user]:[pass]@my-amqp-address',
            amqpQueue: '[my-queue]'
        }
    }).start();

That's it! Now communication from the python process and the node webserver process goes through rabbitmq.

Credits

Aside from the obvious ones:

• The dashboard is built on React-Redux starter kit