TopolearnAutolog gives complete runs with one line, so nothing important is forgotten.
mlflow.autolog() patches supported libraries so that training calls log themselves. For scikit-learn, calling fit creates (or uses) a run and records:
- all estimator parameters from
get_params(), - training metrics such as accuracy, F1, ROC AUC for classifiers or MSE and R2 for regressors,
- the fitted model with an inferred signature,
- for searches like
GridSearchCV, the best parameters and child runs for candidates.
Library-specific versions offer more options, e.g. mlflow.sklearn.autolog(log_models=False, max_tuning_runs=10). Supported integrations include scikit-learn, XGBoost, LightGBM, Keras/TensorFlow, PyTorch Lightning, Spark, statsmodels, and GenAI libraries such as OpenAI and LangChain, where autolog produces traces instead.
Autolog composes with manual logging: inside a with mlflow.start_run(): block, your own log_metric calls join the same run. Evaluation metrics on a test set are not logged automatically for sklearn unless you call a scoring function (MLflow records post-training metrics for calls like model.score or sklearn.metrics functions after fit).
Call autolog() before you import and train, and turn it off with mlflow.autolog(disable=True). It is ideal for exploration; for production pipelines many teams log explicitly so the contents of a run are deliberate.
Example
import mlflow
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
mlflow.sklearn.autolog()
X, y = load_iris(return_X_y=True)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, random_state=0)
with mlflow.start_run() as run:
model = RandomForestClassifier(n_estimators=50, max_depth=3, random_state=0)
model.fit(X_tr, y_tr) # params, train metrics, model logged
model.score(X_te, y_te) # post-training metric logged too
r = mlflow.get_run(run.info.run_id)
print(r.data.params["max_depth"]) # 3
print(sorted(k for k in r.data.metrics if k.startswith("training_"))[:3])
# ['training_accuracy_score', 'training_f1_score', 'training_log_loss']Remember
mlflow.autolog()logs params, metrics and the model for supported libraries- Manual logging inside the same run adds to autologged data
- GenAI library autolog produces traces rather than training metrics
Watch out for
- Enabling autolog after the model was already trained, so nothing gets captured.