Low-code AutoML · train, compare, tune & deploy in a few lines · verified against PyCaret 3.4 (2026)

PyCaret cheat sheet

PyCaret wraps scikit-learn, XGBoost, LightGBM and more behind a tiny, consistent API: setup() builds a preprocessing pipeline, compare_models() trains & ranks a dozen models by cross-validation, and a handful of verbs tune, ensemble, interpret, and deploy the winner. Great for fast baselines and low-code ML across classification, regression, clustering, anomaly detection and time series. This sheet targets the stable PyCaret 3.4 functional API (note: 4.0, in alpha, moves to an OOP-only *Experiment API).

setup compare & create tune & ensemble analyze & predict finalize & deploy gotcha ★ most common

Verified 2026-08-24 against the official docs at pycaret.gitbook.io / pycaret.readthedocs.io and the pycaret/pycaret repo (3.4 stable; 4.0 alpha). Works with scikit-learn 1.x, pandas 2, NumPy 2 on Python 3.9–3.12+.

Outline

The whole workflow is ~6 calls: setup → compare_models → tune_model → predict_model → finalize_model → save_model. Everything else refines a step.

Prepare

setup() does all the preprocessing — it's the most important call.

1Install & setup()3.4

pip install pycaret # pycaret[full] for all optional models/tools★
Heavy install (many ML libs). Consider a fresh venv.
from pycaret.classification import *★
Pick the task module: classification, regression, clustering, anomaly, time_series. Same verbs across all.
s = setup(data=df, target="churn", session_id=123)★
The core call: builds the train/test split + a full sklearn preprocessing pipeline (impute, encode, scale). session_id = reproducibility.
setup(..., train_size=0.8, fold=5, use_gpu=True)
Control split size, CV folds, and GPU (for XGBoost/LightGBM/CatBoost).

2Preprocessing optionssetup args

setup(..., normalize=True, normalize_method="zscore")★
Scaling, plus transformation (power/quantile) for skewed features.
setup(..., categorical_features=[...], numeric_imputation="mean")
Types are auto-inferred but overridable; control imputation, rare-category grouping, encoding.
setup(..., fix_imbalance=True, remove_multicollinearity=True, feature_selection=True, pca=True)★
One-flag SMOTE, collinearity/variance filtering, feature selection, PCA — all folded into the pipeline.
setup(..., log_experiment=True, experiment_name="exp1")
Auto-log everything to MLflow — params, metrics, artifacts per model.

Train

Let PyCaret pick a winner, or train a specific model.

3compare_modelsthe AutoML step

best = compare_models()★
Cross-validates ~15 models and returns the best by the default metric — prints a sorted leaderboard.
top3 = compare_models(n_select=3, sort="AUC")★
Return several models (for ensembling) and rank by any metric.
compare_models(include=["lr","rf","lightgbm"], exclude=["knn"])
Restrict the field to speed it up. budget_time=5 caps the run to N minutes.
pull() # the last results grid as a DataFrame★
Grab any verb's scoring table programmatically. models() lists every available model id.

4create_modela specific model

lgbm = create_model("lightgbm")★
Train one model by id with cross-validation; prints per-fold scores. Ids: lr, rf, xgboost, catboost, et, ...
create_model("rf", fold=10, n_estimators=200)
Pass model kwargs directly; override folds.
create_model("dt", cross_validation=False)
Skip CV for a quick single fit.

Optimize

Tune hyperparameters and combine models.

5tune_modelhyperparameters

tuned = tune_model(lgbm, n_iter=50, optimize="AUC")★
Randomized search over a sensible grid. optimize picks the target metric.
tune_model(lgbm, search_library="optuna", search_algorithm="tpe")★
Swap the backend for Optuna/Hyperopt/scikit-optimize instead of the default random search.
tune_model(lgbm, custom_grid={"n_estimators": [100, 300, 500]})
Provide your own search space; choose_better=True keeps the tuned model only if it actually beats the input.

6Ensemblescombine models

bagged = ensemble_model(dt, method="Bagging") # or "Boosting"
Bag/boost a single estimator.
blender = blend_models([lr, rf, lgbm])★
Soft/hard voting across several models — often a quick accuracy bump over any single one.
stacker = stack_models(top3, meta_model=lr)★
Stacked generalization: a meta-model learns from the base models' predictions.

Understand & Predict

Diagnose the model, then score new data.

7plot / evaluate / interpretdiagnostics

plot_model(tuned, plot="auc") # confusion_matrix, feature, pr, ...★
Dozens of diagnostic plots. plot="feature" = importances, "confusion_matrix", "error".
evaluate_model(tuned)★
Interactive widget (in a notebook) to flip through all plots for a model.
interpret_model(tuned) # SHAP summary★
SHAP-based explanations for tree models — global importance & per-prediction reasons.
dashboard(tuned) · check_fairness(tuned, ...)
An ExplainerDashboard and fairness/bias checks for deeper analysis.

8predict_modelscore data

predict_model(tuned) # on the held-out test set★
Evaluate on the internal hold-out (created by setup) — the honest generalization estimate.
preds = predict_model(tuned, data=new_df)★
Score unseen data — the full preprocessing pipeline is applied automatically. Adds prediction_label/prediction_score columns.

Ship

Retrain on all data, persist, and deploy.

9finalize, save, deployproduction

final = finalize_model(tuned)★
Retrains on the entire dataset (train + hold-out) — do this last, once you've chosen the model.
save_model(final, "churn_pipeline") model = load_model("churn_pipeline")★
Persists the whole pipeline (preprocessing + model) as one artifact. Reload and predict_model anywhere.
deploy_model(final, model_name="m", platform="aws", authentication={...})
Push to S3/GCS/Azure. create_api/create_docker scaffold a FastAPI service + Dockerfile.
get_config("X_train") · set_config(...)
Reach into the experiment's internals (data splits, pipeline) when you need to.

!Common gotchasread before shipping

setup() defines everything
Preprocessing, the split, and CV are all fixed at setup time. Re-run setup to change them — downstream verbs inherit it.
finalize AFTER model selection
finalize_model trains on the hold-out too, so its scores are no longer an unbiased estimate. Choose the model first, finalize last.
save_model saves the pipeline
Don't pickle the raw estimator — use save_model so preprocessing travels with it, or new-data predictions will be wrong.
low-code ≠ no-thought
compare_models gives a baseline, not a final answer. Check for leakage, validate the metric, and understand the winner before shipping.
3.x functional vs 4.0 OOP
4.0 (alpha) removes the functional API for ClassificationExperiment-style objects. This sheet is stable 3.x — verify before adopting 4.0.

Worth memorizing

setup → compare → tune → predict → finalize → savethe whole workflow in six calls

same verbs, five tasksimport pycaret.classification / regression / clustering / anomaly / time_series

setup() is the pipelinesplit + impute/encode/scale + options like fix_imbalance, feature_selection

compare_models = AutoML leaderboardn_select for several; pull() for the grid

tune_model with optuna backendchoose_better keeps it only if it wins

blend/stack for a bumpcombine the top models

interpret_model = SHAPplot_model/evaluate_model for diagnostics

finalize last, then save the pipelinesave_model bundles preprocessing + model

log_experiment=Trueauto MLflow tracking

PyCaret cheat sheet (pycaret_cheatsheet v0).
Verified 2026-08-24 against pycaret.gitbook.io / pycaret.readthedocs.io and pycaret/pycaret. Stable 3.4 functional API (from pycaret.<task> import *) over scikit-learn/XGBoost/LightGBM/CatBoost; MLflow tracking built in. 4.0 (alpha) shifts to an OOP-only *Experiment API — confirm the version before adopting. Low-code gives fast baselines, not a substitute for validation & leakage checks.
Design system = shared cheat-sheet tokens (Inter + JetBrains Mono, light theme). Content only.