Course contentsShow
← Machine Learning and Deep Learning
Lesson 2814 of 3,53861. Experiment Tracking and Model ManagementFree lesson

What to Track in ML Experiments

Core elements to log: hyperparameters, metrics, code version, data snapshots, environment, random seeds, and model artifacts.

What to Track in ML Experiments

What you'll learn: The essential elements you must log during machine learning experiments to ensure reproducibility and enable systematic improvement.

The Six Pillars of Experiment Tracking

When you run an ML experiment, you're not just training a model—you're generating evidence that must be preserved. Think of each experiment as a lab notebook entry: without complete notes, you can't reproduce your success or learn from failures.

Core Elements to Log

1. Hyperparameters
Every knob you turn: learning rate, batch size, number of layers, dropout rates, optimizer choice. These are your independent variables.

2. Metrics
Training loss, validation accuracy, F1 scores, custom business metrics—tracked over time, not just final values. These are your dependent variables.

3. Code Version
Git commit hash or version tag. Your exact code state at experiment time. Without this, you're flying blind even if everything else is logged.

4. Data Snapshots
Dataset version, preprocessing steps, train/val/test split seeds. The data fingerprint ensures you know exactly what went into your model.

5. Environment
Library versions (PyTorch 2.0.1, CUDA 11.8), hardware specs, OS details. A different CUDA version can silently change results.

6. Random Seeds
All seeds: NumPy, PyTorch, Python's random module. The difference between reproducible science and frustrating mystery.

7. Model Artifacts
Saved checkpoints, final weights, architecture diagrams. The physical outputs of your experiment.

The Analogy

Imagine debugging a recipe where you forgot whether you used butter or oil, what oven temperature you set, and which brand of flour you bought. That's ML without proper tracking.

Key Takeaway: Comprehensive experiment tracking—hyperparameters, metrics, code version, data snapshots, environment, seeds, and artifacts—transforms chaotic trial-and-error into systematic, reproducible science.