Autologging, search and the UI

Capturing runs with one line, organising hyperparameter searches, and finding the best run in code or in the UI.

Autologging

Also called mlflow.autolog

What: mlflow.autolog() hooks into supported libraries such as scikit-learn, XGBoost, LightGBM, PyTorch Lightning and Keras, and logs parameters, metrics and the model when you call fit.
Why: Complete runs with no logging code, which is ideal for notebooks and quick comparisons.

Python

import mlflow
from sklearn.ensemble import RandomForestClassifier

mlflow.autolog()
with mlflow.start_run():
    clf = RandomForestClassifier(n_estimators=200)
    clf.fit(X_train, y_train)

Watch out

Enabling autolog and also logging the same model by hand, producing two copies per run.

Library-specific autolog

What: Each integration also has its own autolog, such as mlflow.sklearn.autolog(), with options like log_models, log_datasets and disable.
Why: Fine control: log metrics automatically but register models by hand, or silence one library.

Python

mlflow.sklearn.autolog(log_models=False, log_datasets=False)
mlflow.xgboost.autolog()

Nested runs

Also called parent and child runs

What: Runs started with nested=True inside another run become its children, and the UI groups them under the parent.
Why: A hyperparameter search reads as one parent with a child per trial instead of hundreds of loose runs.

Python

with mlflow.start_run(run_name="depth-search"):
    for depth in [4, 8, 16]:
        with mlflow.start_run(run_name=f"depth={depth}", nested=True):
            mlflow.log_param("max_depth", depth)
            mlflow.log_metric("val_auc", train_and_score(depth))

Searching runs

Also called search_runs, filter_string

What: mlflow.search_runs returns a pandas DataFrame of runs matching a SQL-like filter on metrics, params, tags and attributes, in the order you ask.
Why: Pick the best run in code, for example as a pipeline step, instead of clicking through the UI.

Python

runs = mlflow.search_runs(
    experiment_names=["churn-model"],
    filter_string="metrics.val_auc > 0.85 and params.max_depth = '8'",
    order_by=["metrics.val_auc DESC"],
    max_results=5,
)
print(runs[["run_id", "metrics.val_auc"]])

Watch out

Comparing params as numbers; they are stored as strings, so quote them in filters.

MlflowClient

What: The lower-level API for reading and editing runs, experiments, metric histories and registered models by id.
Why: Scripts and CI jobs use it to inspect runs and manage the registry without an active run.

Python

from mlflow import MlflowClient

client = MlflowClient()
run = client.get_run(run_id)
print(run.data.metrics, run.data.params)
history = client.get_metric_history(run_id, "val_loss")
client.set_tag(run_id, "reviewed", "true")

The MLflow UI

What: A web app listing experiments and runs, with metric charts, side-by-side run comparison, parallel coordinates plots, artefact browsing, models and traces.
Why: Comparing a dozen runs visually is often the fastest way to see which parameter actually matters.

Shell

mlflow ui --port 5000

Organising runs

What: Clear experiment names, meaningful run names and consistent tags, plus descriptions on experiments explaining the question being explored.
Why: Six months later, "churn-model / xgb-new-features" beats "Default / dapper-owl-42".

Go deeper