MLbot - Jupyter notebook version control, experiment tracking

Installation

pip intall mlbot

The project is under constant development of alpha version. pip install --upgrade mlbot to upgrade to newest version.

Important

After installation/upgrade, restart jupyter notebook instace (restart cloud instance if using one) for the changes to take effect.

Features

1. Version control

   • multiple storage options
   • show historical versions in dropdown menu
   • switch to specific version of the notebook with a click

2. Automatically experiment tracking

   • parameters, loss, metrics, data version, model etc....
   • using MLFlow at the moment:
     • by default logging locally to current working directory
       • run mlflow ui in the current working directory terminal
       • tracking ui is then served at localhost:5000/
     • if MLFLOW_TRACKING_URI and AWS credentials are set up, logging to remote server

Todo:

• save to github
• save to gitlab

Limitations:

• when the notebook is renamed, it'll generate a new save folder for the newly named notebook without access to previous versions. Can be possibly solved by creating custom ContentManager API.
• MLFlow automatic experiment tracking only works with fastai at the moment. Support for other languages coming later.