How to Forecast Through MLflow's Generic Predict Interface¶

This guide shows you how to forecast with a yohou model loaded through mlflow.pyfunc.load_model. Use it when your code loads MLflow models generically rather than through yohou_mlflow.load_model.

Prerequisites¶

A forecaster saved or logged with yohou-mlflow
New observations as polars DataFrames, if you want to forecast from them

Steps¶

1. Load the model¶

import mlflow.pyfunc

model = mlflow.pyfunc.load_model("models:/daily-demand/latest")

Load only models you trust this way: MLflow imports the modules the model file names before any check runs. For a model you have not reviewed, run check_compatibility first, or use yohou_mlflow.load_model.

If the forecaster contains your own estimator classes, pass them at load time: mlflow.pyfunc.load_model(uri, model_config={"extra_trusted_types": [...]}).

2. Forecast¶

To forecast from the saved state, pass an empty dict:

forecast = model.predict({})

To forecast from new observations, pass them under y. The model observes them and forecasts once, from the last observed time:

forecast = model.predict({"y": new_rows})

Pass exogenous features under X_actual, X_future and X_forecast, the names yohou uses. X_actual is only accepted together with y:

forecast = model.predict({"y": new_rows, "X_actual": new_features, "X_future": known_future})

3. Choose what to predict¶

Pass params to select the prediction and its options:

Prediction intervals:

forecast = model.predict({"y": new_rows}, params={"prediction_type": "interval", "coverage_rates": [0.8, 0.95]})

Class probabilities, for a class-probability forecaster:

forecast = model.predict({"y": new_rows}, params={"prediction_type": "class_proba"})

A different number of steps, or some panel groups only:

forecast = model.predict({}, params={"forecasting_horizon": 3, "groups": ["store_1"]})

The result is a polars DataFrame, the same frame the yohou method returns.

4. Keep the update, if you need it¶

Each predict call works on a copy, so the observations are not kept. To keep them, load the forecaster itself, observe, and register it again:

import yohou_mlflow

forecaster = yohou_mlflow.load_model("models:/daily-demand/latest")
forecaster.observe(new_rows)
with mlflow.start_run():
    yohou_mlflow.log_model(forecaster, name="forecaster", registered_model_name="daily-demand")

Troubleshooting¶

Problem: TypeError: The yohou pyfunc input must be a dict of polars DataFrames: Pass a dict of polars frames. pandas input, and therefore mlflow models serve and Spark UDFs, is not supported yet.
Problem: prediction_type='interval' is not supported by this forecaster: The forecaster cannot produce that prediction. The error lists the types it supports.