MLflow

MLflow documenation Example Usage

What it is¶

MLflow is a local experiment tracker. It logs parameters, metrics, and artifacts for every training run and provides a web UI to compare them side-by-side. No account or internet connection is required as everything lives in a tracking directory on disk.

Why use it¶

Without a tracker, comparing runs means maintaining spreadsheets or reading output filenames. MLflow replaces that with a queryable log of every run which includes parameters, metrics, plots, and model artifacts plus a local UI you can open in a browser.

It also ships with a local Model Registry, so you can version named models and load a specific one by alias later.

When to use it¶

• Solo or local work where a cloud account is unnecessary.

• Offline HPC environments where outbound network access is restricted.

• When you want collaborators to view runs without creating cloud accounts.

• When you want a local model registry to load specific model versions programmatically.

How to use it¶

Install¶

uv add mlflow

Configure with environment variables¶

export MLFLOW_EXPERIMENT_NAME="sst_enso_experiment"
export MLFLOW_RUN_NAME="baseline"
export MLFLOW_TRACKING_DIR="runs/sst_enso/mlruns"
export DATA_ROOT="./data"

Run and track¶

source .venv/bin/activate
python scripts/train_sst_mlflow.py --run-name "baseline"

Launch the UI¶

Point mlflow ui at the directory that contains your tracking data. Two forms come up in practice:

# Point directly at the tracking directory
mlflow ui --backend-store-uri runs/sst_enso/mlruns

# Or point at a parent that holds multiple experiments
mlflow ui --backend-store-uri /path/to/runs

Open the printed URL (default http://127.0.0.1:5000) to browse runs.

What gets tracked per run¶

Each run in workshop-sst logs:

The DVC MD5 tags are the key to reproducing a past run — see MLflow + DVC.

Reload and re-run a past model¶

MLflow’s Model Registry makes it easy to load a specific version without hunting through files:

import mlflow.sklearn

# Latest version of the registered model
model = mlflow.sklearn.load_model("models:/sst_enso_predictor/latest")

# A specific version
model = mlflow.sklearn.load_model("models:/sst_enso_predictor/3")

# Using an alias (e.g. "staging", "production")
model = mlflow.sklearn.load_model("models:/sst_enso_predictor@staging")

predictions = model.predict(X_test)

If the model was logged but not registered, you can still load it by run ID:

model = mlflow.sklearn.load_model("runs:/<run_id>/model")

Share runs with a collaborator¶

MLflow stores everything in the tracking directory, so sharing is mostly a question of sharing that directory.

# Option 1: zip the tracking directory and hand it off
zip -r sst_runs.zip runs/sst_enso/mlruns/

# Option 2: put the tracking store on shared or cloud storage
export MLFLOW_TRACKING_URI="s3://your-bucket/mlruns"
# or an NFS path:
export MLFLOW_TRACKING_URI="file:/mnt/shared/mlruns"

For option 2, every collaborator pointing at the same URI sees the same runs live.

Pros and cons¶

Reference¶

• MLflow documentation

• MLflow Tracking

• MLflow Model Registry

• Next: pair MLflow with DVC for full reproducibility in MLflow + DVC.