Machine Learning for Software Engineering: modelling the source code

Slides are online.

OSS tools covered:

Abstract

Machine Learning on Source Code (MLonCode) is an emerging research domain which stands at the intersection of deep learning, natural language processing, software engineering and programming language communities.

During this 3h30 workshop, we will review recent Software Engineering tasks that benefit from applying Machine Learning, with a focus on hands-on experience on:

extracting data from real source code

developing multiple Machine Learning models

for a particular task of source code summarization (or function name suggestion).

At the end of the workshop participants will build 2 working models on a real dataset, producing near state-of-the-art results. Practical skill of extracting information from source code as well as modelling different aspects of it are going to be acquired.

Prerequisites: familiarity with the basics of DeepLearning, a laptop with Docker installed

Prerequisites

Docker

Dependencies

Import Docker images (works offline):

docker load -i images/jupyter.tgz
docker load -i images/gitbase.tgz
docker load -i images/bblfshd-with-drivers.tgz

docker images

Run bblfsh

docker run \
    --detach \
    --rm \
    --name amld_bblfshd \
    --privileged \
    --publish 9432:9432 \
    bblfsh/bblfshd:v2.15.0-drivers \
    --log-level DEBUG

Run gitbase

docker run \
    --detach \
    --rm \
    --name amld_gitbase \
    --publish 3306:3306 \
    --link amld_bblfshd:amld_bblfshd \
    --env BBLFSH_ENDPOINT=amld_bblfshd:9432 \
    --env MAX_MEMORY=1024 \
    --volume $(pwd)/repos/git-data:/opt/repos \
    srcd/gitbase:v0.24.0-rc2

Run the jupyter image

docker run \
    --rm \
    --name amld_jupyter \
    --publish 8888:8888 \
    --link amld_bblfshd:amld_bblfshd \
    --link amld_gitbase:amld_gitbase \
    --volume $(pwd)/notebooks:/amld/notebooks \
    --volume $(pwd)/repos:/amld/repos \
    mloncode/amld

With make

To build the workshop image and launch the 3 required containers

make build-and-run

To only launch the 3 required containers

make

Workflow

1. Download the data

We are going to use top 50 repositories from Apache Software Foundation though this workshop.

Notebook 1: data collection pipeline (example)

2. Project and Developer Similarities

Build a vector model for projects and developers using Topic Modelling of code identifiers.

Notebook 2: project and developer similarities (example)

3. Function Name Suggestion

Train a NMT seq2seq model for predicting method names based on identifiers in method bodies.

Notebook 2: function name suggestion (example)