Deploy

Deployment in PyCaret 4.0 takes one of three shapes:

1. Just persist — save_model to disk, load_model later.
2. Self-host — drop the saved Pipeline behind a FastAPI route.
3. Use pycaret-server — the bundled deployment service that ships with the dashboard.

finalize_model(estimator)#

Refits the input estimator on the full dataset (train + test combined). Use it as the last step before deploying — the holdout has already done its job, so squeeze it back into training.

final = exp.finalize_model(tuned.pipeline)
final.pipeline   # sklearn Pipeline fit on all available data

The input is left untouched — finalize_model returns a fresh Pipeline. If you call predict on the original after finalizing, you'll still get the train-only fit's predictions.

For time-series, "full" means y_train + y_test concatenated. The finalized model can then forecast fh periods past the actual end-of-data.

save_model(estimator, name)#

Joblib-pickles the Pipeline to <name>.pkl in the current directory:

exp.save_model(final.pipeline, "production-juice-classifier")
# → production-juice-classifier.pkl

The file contains the full sklearn / sktime Pipeline — preprocessor

• trained model — so anyone with sklearn installed can joblib.load it and call .predict(X).

load_model(name)#

Loads a previously-saved Pipeline:

from pycaret.classification import ClassificationExperiment
exp = ClassificationExperiment(target="Purchase")
pipeline = exp.load_model("production-juice-classifier")
predictions = pipeline.predict(new_data)

You don't need the original Experiment to use a saved Pipeline — it's a real sklearn Pipeline. This works:

import joblib
pipeline = joblib.load("production-juice-classifier.pkl")
predictions = pipeline.predict(new_data)

Self-host behind FastAPI#

The simplest production deployment is twenty lines:

from fastapi import FastAPI
import pandas as pd
from joblib import load

pipeline = load("production-juice-classifier.pkl")
app = FastAPI()

@app.post("/predict")
def predict(rows: list[dict]):
    df = pd.DataFrame(rows)
    return {"predictions": pipeline.predict(df).tolist()}

Run it:

uvicorn server:app --host 0.0.0.0 --port 8000

That's it. No PyCaret-specific deployment tooling required.

pycaret-server#

The optional pycaret-server (in services/api/) wraps every saved Pipeline behind a managed deployment with:

• Workspace-scoped pipeline registry — promote a fitted Pipeline to a named registry entry.
• HTTP endpoint — /api/v1/deployments/{slug}/predict accepts rows and returns predictions, with rolling p50 / p95 latency tracking.
• Audit log — every prediction call is recorded.
• Drift monitoring — periodic comparison of incoming feature distributions against the training baseline.

Install it:

pip install "pycaret[dashboard]"
pycaret-server          # starts the FastAPI app

Then use the dashboard or the HTTP API to promote and deploy.

What's removed#

The kill list philosophy: PyCaret should help you build a model. Your deployment infrastructure (cloud, container, mobile, edge) is your infrastructure. We don't try to be all of those at once anymore.