A lightweight approach to experiment tracking

When you are training ML models it's a good idea to keep track of what you have done. I've been looking into a few experiments tracking solutions (with the help of the Matt at FuzzyLabs). Some of the ones I looked at were too heavyweight and required changes in the way you write your training code. I wanted something that just works out of the box.

What I found was guild.ai. If your trying script is train.py, it's as simple as replacing:

python train.py

with

guild run train.py

With one command, you automatically record:

• Experiment inputs
• Experiment results
• Snapshot of source code

The runs get stored using your normal file system - no need to maintain a separate database. You can then go back and look at previous runs using:

guild view