/github-contributions-predictor

How predictable am I in my GitHub contributions

Primary LanguageJupyter NotebookGNU Affero General Public License v3.0AGPL-3.0

GitHub's contributions Predictor

A toy project to see how predictable I'm with my so-called GitHub contributions ;)

One of the main goal of this repo is to predict current/next day contributions of multiples users in a daily automated way using GitHub actions.
To do so this project feature a pytorch model trained with contributions data from GitHub users.
The history of those predictions is available in the pred_history_no_scaling branch

How to get predictions for your contributions

Please first consider that this project is just for fun, not well tested and intended for an harmless use.

To add a user for the next predictions, do the following:

  • fork this repository
  • append your GitHub nickname into the users.txt file
  • commit
  • open a pull request

Technical Process Overview

The following readme parts are now more technical.

Here's an overview of the process to predict contributions from zero:

  1. Gather contributions data
  2. Train a machine learning model
  3. Use the model to predict futures contributions (published here)
  4. Repeat 3. every day by using GitHub actions

Requirements to rebuild a model

  • anaconda
  • pytorch (with or without GPU)
  • any additional pip requirements are listed in requirements.txt

Source files description

To allow one to build his own model the project is organized in multiple ordered python/jupyter files designed to be ran sequentially.

0-gather_data.py

Download and save users' contributions and other stats provided by GitHub public api.
User list is collected by randomly walking the users' following/followers graph.
Produce a big contribs.json files containing raw users' data.
This script can be run again to gather even more data.

1-pack-data.py

Parse and pack gathered data into numpy ndarrays.
Produce a compressed userdata.npz numpy file

2-preprocess.py

Pre-process users' contributions by using the following scheme:

  • data augmentations using mean, std, skewness and fft
  • outliers removal using quantiles filters mainly
  • features normalization using scikit-learn preprocessing tools

Produce a compressed ml.npz numpy file and a scalers.pkl.z containing pickled scalers.

3-train-model.ipynb

Jupyter notebook (designed to be run on kaggle) for training a pytorch model.

4-inference.py

Use previous pytorch model, download the latest users' data and predict their contributions number for the next 7 days.
Produce csv files containing predictions.

Bonus for the readers

There's an additional branch pred_history_with_scaling containing predictions with a model trained to expect more contributions from users.