AutoML Experiments

A collection of AutoML experiments than can be executed in Docker and can use Kafka as streaming data source.

Running instructions

Requirements

Required: Docker

Strongly recommended: Docker Compose, Make

Useful: kafkacat

Starting the containers

All containers at once:

make up

Individual containers:

docker-compose up auto-sklearn zookeeper broker

Publishing a dataset to Kafka

OpenML dataset:

make publish-openml-dataset

For any other dataset:

cat ./datasets/covtype.csv | kafkacat -P -b localhost -t covtype   

Running an experiment

make train-scikit-multiflow-kafka

Or directly using Docker Compose

docker-compose exec auto-sklearn python training/scikit-multiflow-kafka.py

Alternatively, you can run a single container using only Docker run.

Opening Jupyter/JupyterLab

Find the right port for the experiment/service in the docker-compose.yml

Navigate to: localhost:<port>, for example: localhost:8888

Get the Jupyter token by running

docker-compose logs <service_name>

For example:

docker-compose logs auto-sklearn

Copy the token and use it to login in Jupyter.

Stopping the containers

All containers at once:

make down

Development instructions

For developing the experiments it is useful to have the dependencies installed locally in a virtualenv. It helps IDEs to provide autocompletion information.

• Create and activate a virtualenv
• Install some or all dependencies from dev-requirements.txt

pip install -r dev-requirements.txt