Autologging

mlflow_autogluon.autolog() patches the fit method of every installed AutoGluon predictor type (TabularPredictor, TimeSeriesPredictor, MultiModalPredictor) using MLflow's own autologging machinery (safe_patch), so it behaves like the built-in integrations: runs are created automatically when none is active, user-created runs are reused, and a logging failure never breaks your training call.

Options

mlflow_autogluon.autolog(
    log_models=True,             # log the fitted predictor as an MLflow model
    log_model_signatures=True,   # infer a signature from training data + predictions
    log_input_examples=False,    # save a training-feature sample with the model
    log_datasets=True,           # attach the training data as an MLflow dataset input
    log_leaderboard=True,        # leaderboard CSV artifact + per-model metrics
    log_fit_summary=False,       # fit_summary() as a JSON artifact
    registered_model_name=None,  # also register logged models under this name
    extra_tags=None,             # extra tags applied to autologged runs
    disable=False,               # turn the integration off
    silent=False,                # suppress autologging warnings
)

What gets logged

Parameters

All non-data arguments of the fit call, plus predictor configuration:

label, problem_type, eval_metric from the predictor (target and prediction_length for timeseries predictors)
presets, time_limit, hyperparameters, num_bag_folds, num_stack_levels, and any other keyword arguments passed to fit
train_rows and train_columns when the training data has a shape

Dict and list values (such as hyperparameters) are JSON-encoded and truncated to 500 characters to stay within MLflow param limits.

Metrics

fit_time_seconds: wall-clock duration of the fit call
best_model_score_val: validation score of the best model
every numeric leaderboard column per model, e.g. score_val_<model>, fit_time_<model>, pred_time_val_<model> (column sets vary by predictor type)

MultiModalPredictor has no leaderboard, so leaderboard metrics and the CSV artifact are skipped for it.

Artifacts

leaderboard.csv: the full leaderboard
fit_summary.json when log_fit_summary=True
model/: the fitted predictor logged with the autogluon flavor (disable with log_models=False)

Signatures, input examples, and datasets

Matching MLflow's built-in autologging conventions:

Model signature (log_model_signatures, default on): inferred from a small sample of the training features and the predictor's output, so logged models enforce their input schema at serving time. For timeseries predictors the signature reflects the long-format (item_id, timestamp, target) contract and the forecast columns. The predict_method inference param is always preserved.
Input example (log_input_examples, default off): saves the sampled training features alongside the model.
Dataset lineage (log_datasets, default on): the training DataFrame is attached to the run as an MLflow dataset input (visible in the run's "Datasets used" panel). File-path training inputs are skipped.

All of this is best effort: a failure in signature inference or dataset logging is downgraded to a warning and never interrupts training.

Registering models automatically

mlflow_autogluon.autolog(registered_model_name="churn-classifier")

Every subsequent fit logs the predictor and creates a new version of churn-classifier in the model registry.

Using your own runs

Autologging respects an already-active run:

import mlflow

with mlflow.start_run(run_name="experiment-42"):
    predictor.fit(train_data)          # logs into experiment-42
    mlflow.log_param("my_param", 1)    # your own logging works alongside

Warning

Calling fit twice inside the same run logs parameters twice. MLflow rejects conflicting param values, and mlflow-autogluon downgrades that rejection to a warning so training is never interrupted; the second fit's params are simply not recorded. Use one run per fit when you need full param capture.

Disabling

mlflow_autogluon.autolog(disable=True)