AMF: Aggregated Mondrian Forests for Online Learning

This GitHub page is a helper to reproduce the experiments proposed in the paper

AMF: Aggregated Mondrian Forests for Online Learning

by J. Mourtada, S. Gaïffas and E. Scornet

arXiv link: http://arxiv.org/abs/1906.10529

The AMF algorithm is implemented as the class OnlineForestClassifier in the tick.online module of the tick library https://github.com/X-DataInitiative/tick. This module is not yet merged into the master branch of tick and therefore is not yet easily installable via pip. It must be compiled and installed from source for now. Therefore, this GitHub repository is here to ease this installation, through the use of a Docker image.

1. Installation

First, you need to install Docker on your computer https://www.docker.com/get-started, and you need to create an account on https://hub.docker.com/ in order to use the image I created for you. It will make the installation of all the tools required to run the experiments 1000 times easier, and it won't mess with your local configuration files and Python environments. Then, you need to git clone this repository somewhere by typing

git clone https://github.com/stephanegaiffas/AMF.git

in a terminal. Now, you need to run the image that contains everything you need. This image is called stephanegaiffas/amf:v1. Once Docker is installed, you can simply type in a terminal

docker run -it -v <PATH>/AMF/:/AMF stephanegaiffas/amf:v1

where <PATH> is the path leading to the AMF directory corresponding to this repository. On my laptop it is /Users/stephanegaiffas/Code. You must change only the <PATH> in this command line and nothing else, note that stephanegaiffas/amf:v1 is the name of the image I built for you on docker-hub.

This command-line runs the image and accesses the container (instance of the image) in interactive mode. This means that after running this command-line, you are inside the container, where everything has been already installed for you. Note that Docker will download automatically the image the first time you launch this command-line, which may take some time (the download is around 2GB).

2. Run the experiments

Once inside the container you can simply run the scripts to reproduce the experiments of the paper. For instance

python3 plot_decisions.py

will create a decisions.pdf file, which corresponds to the second figure of the paper. Note that in this image Python is called using the python3 command.

[ ] Detail all the scripts and what they do

3. Explain the docker-serve experiment

[ ] TODO

Appendix. Build the docker image from the Dockerfile

Warning: only for those who know what they do and need to modify the image

The Dockerfile can be used to build the image containing all the tools required (although a pre-built image is available on docker hub, see above). In order to generate this image, simply run

docker build -t amf:v1 .

in the folder containing the Dockerfile. This will take a very long time (almost one hour !) since it configures everything from a basic Linux image, and since the compilation of tick is very long. This image can be now shared through a

docker tag amf:v1 stephanegaiffas/amf:v1
docker push stephanegaiffas/amf:v1

(although with another user name...)