Quickstart

Install

pip install mlflow-autogluon[tabular]

The tabular extra pulls in autogluon.tabular. If AutoGluon is already installed, plain pip install mlflow-autogluon is enough.

Autologging in three lines

import mlflow_autogluon
from autogluon.tabular import TabularPredictor

mlflow_autogluon.autolog()

predictor = TabularPredictor(label="target").fit(train_data)

That is the whole integration. Every fit call creates (or reuses) an MLflow run and records:

What	Examples
Params	label, problem_type, eval_metric, presets, time_limit, hyperparameters, train_rows
Metrics	best_model_score_val, fit_time_seconds, per-model score_val_* / fit_time_*
Tags	estimator_name, autogluon_version, best_model, problem_type
Artifacts	leaderboard.csv, optional fit_summary.json, the fitted model

Note

Because this is a community flavor, mlflow.autolog() does not enable it automatically. Call mlflow_autogluon.autolog() explicitly.

Load the model back

import mlflow
import mlflow_autogluon

run_id = mlflow.last_active_run().info.run_id

# As the native AutoGluon predictor
predictor = mlflow_autogluon.load_model(f"runs:/{run_id}/model")

# Or as a generic pyfunc
pyfunc_model = mlflow.pyfunc.load_model(f"runs:/{run_id}/model")
predictions = pyfunc_model.predict(test_data)

Where to go next

• Autologging for all configuration options
• Model flavor for manual log_model / save_model workflows
• Serving for REST scoring with mlflow models serve