ml-experiment-tracking

ML Experiment Tracking

When to Use

When training models and need to compare runs, reproduce results, or share findings with the team.

Core Jobs

1. Track Every Run

Use MLflow, W&B, or DVC. Log per experiment:

• Parameters: hyperparameters, data version, preprocessing config
• Metrics: loss, accuracy, precision/recall/F1, AUC
• Artifacts: model weights, feature importance plots, confusion matrix
• Environment: Python version, library versions, hardware

import mlflow

with mlflow.start_run():
    mlflow.log_param("learning_rate", 0.01)
    mlflow.log_param("max_depth", 5)
    mlflow.log_metric("val_accuracy", 0.92)
    mlflow.log_artifact("model.pkl")

2. Version Your Data

• Pin dataset version alongside each experiment
• Use DVC or S3 versioning
• Never train on "latest" — always a specific version

3. Make Experiments Reproducible

• Set random seeds (torch.manual_seed, np.random.seed)
• Record hardware (GPU model affects float precision)
• Use Docker or conda envs with pinned versions
• Log git commit hash with each run

4. Compare and Select

• Don't compare metrics in isolation — look at training curves
• Check for overfitting (train vs val gap)
• Compare on holdout test set only once (before that: val set)
• Document why you chose the winner

Key Outputs

• Experiment tracking setup (MLflow/W&B config)
• Run comparison report
• Model selection rationale
• Reproducibility checklist

Anti-Patterns

• Training without logging parameters
• Comparing models trained on different data versions
• Evaluating on test set during model selection (data leakage)
• No seed → can't reproduce results