Initialize

Experiment(...).fit(data)#

The 3.x setup() function is gone. Every preprocessing knob now lives on the Experiment constructor, and fit(data) actually runs the pipeline.

from pycaret.classification import ClassificationExperiment

exp = ClassificationExperiment(
    target="Purchase",
    session_id=42,
    train_size=0.7,
    fold=10,
    normalize=True,
    transformation=True,
).fit(data)

After fit(), the experiment is ready for every other verb. State is stored on exp._fit_state (see Data preparation).

What fit() does#

1. Coerces input — DataFrame for tabular tasks; Series for univariate time-series.
2. Builds the preprocessing pipeline — imputer + encoder (always); plus optional power transform, scaler, outlier filter, feature selector.
3. Splits train / test — stratified for classification, plain for regression, temporal for time-series, none for clustering / anomaly.
4. Builds the fold generator — StratifiedKFold (clf), KFold (reg), or sktime ExpandingWindowSplitter (TS).
5. Builds the model + metric registries — task-specific dicts of estimators and scorers.
6. Stores everything in exp._fit_state.

After this, calling create_model("rf") reads from the registry, fits an estimator on the (transformed) train set, runs CV via the fold generator, and returns a CreateResult.

Return value#

fit() returns self, so the typical pattern is:

exp = ClassificationExperiment(target="Purchase").fit(data)

You can also call it later if you want to instantiate first and fit on a different dataset, or compose with an environment:

exp = ClassificationExperiment(target="Purchase", session_id=42)
# … some other setup …
exp = exp.fit(data)

sklearn-compatible#

Experiment is a sklearn BaseEstimator. After fit(), sklearn's __sklearn_is_fitted__ returns True; tools that introspect sklearn estimators (joblib, model registries, deployment frameworks) will treat your experiment as fitted.

Different signatures for different tasks#

The constructor signature varies by task:

# Tabular supervised: target is required, train_size + fold matter.
ClassificationExperiment(target=..., train_size=..., fold=...).fit(data)
RegressionExperiment(target=..., train_size=..., fold=...).fit(data)

# Tabular unsupervised: no target, no train/test split.
ClusteringExperiment().fit(data)
AnomalyExperiment().fit(data)

# Time-series: fh + seasonal_period instead of target.
TimeSeriesExperiment(fh=12, seasonal_period=12).fit(univariate_series)

What's removed#

See All setup parameters for the full constructor reference.