Skip to content
Deadline to Die

Phase 2: Tracking

Phase 2: Experiment & Data Tracking

In this phase, we learn how to record every aspect of an ML run so it can be audited and reproduced.

🟢 Level 1: MLflow Tracking

MLflow is the industry standard for experiment management.

1. The Metadata Store

  • Parameters: Hyperparameters (Learning rate, Batch size).
  • Metrics: Evaluation numbers (Accuracy, F1, Loss).
  • Artifacts: The saved model file (.pkl, .h5), plots, and log files.
import mlflow

with mlflow.start_run():
    mlflow.log_param("optimizer", "adam")
    mlflow.log_metric("val_acc", 0.94)
    mlflow.log_artifact("plots/confusion_matrix.png")

🟡 Level 2: DVC (Data Version Control)

Git cannot handle 100GB datasets. DVC handles this by versioning the “Pointer” to the data.

2. The DVC Workflow

  1. dvc add data.csv -> Creates data.csv.dvc.
  2. git add data.csv.dvc .gitignore -> Commit the pointer to Git.
  3. dvc push -> Upload raw data to S3/GCS.
  4. dvc pull -> Other developers download the exact version of the data.

🔴 Level 3: Experiment Pipelines

Instead of one giant script, break training into Steps: Download Data -> Preprocess -> Train -> Evaluate.

3. DVC DAGs

DVC can track the dependencies between steps. If the “Train” code hasn’t changed but the “Data” has, DVC knows to re-run the pipeline.