AL

A project to automatically generate supervised learning programs by learning from previously written programs.

We collected approximately 500 programs, extracted dynamic traces, and build a model to predict the next API call based on previous j calls and features of the current data state.

We previously provided a docker demo, but have now moved on to a simple API for use. This is meant for demonstration purposes.

Build

We recommend you create a virtual environment, for example using conda.

conda create -n al-env python=3.6

You can then install the requirements for use of AL.

pushd src/core
pip install -r requirements.txt
popd

Docker container

Some people have had issues using the install suggested above, particularly they encounter issues relating to xgboost. To address this, users can instead build a docker container for AL. We refer the user to https://www.docker.com/ for information on how to install docker.

Once docker is installed, the container can be built by running

docker build . -t al-container

And it can be launched by running

docker run -it -rm al-container

which launches a (temporary) container with bash as an entry point. You can then launch the python interpreter and import AL following the steps in the next subsection.

Using AL

You can import al from src/core for use.

pushd src/core
python

import al
import sklearn.datasets
X, y = sklearn.datasets.make_classification()

m = al.AL()
m.fit(X, y)
progs = m.get_programs()
print(progs[0].pipeline_code())

Note that .pipeline_code prints out a pipeline without some boilerplate (such as splitting X, y into training and validation).