pyfunc Contract¶

Description of the python_function flavour of a yohou model: how it loads, what predict accepts, and what it returns.

Loading¶

model = mlflow.pyfunc.load_model(model_uri, model_config=None)

`model_config` key	Type	Default	Meaning
`extra_trusted_types`	list of string	`[]`	Type names to trust in addition to the trust policy
`extra_trusted_prefixes`	list of string	`[]`	Module prefixes to trust, each ending in `.`
`extra_version_rules`	mapping of string to string	`{}`	Version rules for packages outside the built-in ones
`strict`	bool	`true`	Raise on a package version mismatch; `false` warns and loads

MLflow drops any other key with a warning, including a key the model was not saved with: a model saved by a yohou-mlflow release without the two extra_* keys above accepts only extra_trusted_types and strict. The loading checks are those of load_model; see Saved model format.

Before those checks run, MLflow imports the loader_module and any code directory named in the model's flavors.python_function. Only models you trust should be loaded this way; check others with check_compatibility first.

`predict(model_input, params=None)`¶

Input¶

model_input is a dict. Keys are a subset of yohou_mlflow.INPUT_KEYS; values are polars DataFrames, passed to yohou unchanged.

Key	yohou argument	Required
`y`	`y`	No
`X_actual`	`X_actual`	Only with `y`
`X_future`	`X_future`	No
`X_forecast`	`X_forecast`	No

An empty dict predicts from the saved state.

Params¶

The signature declares these params (yohou_mlflow.PARAM_NAMES):

Param	Type	Default	Meaning
`prediction_type`	string	`default_prediction_type` from `MLmodel`	One of `point`, `interval`, `class_proba` (`yohou_mlflow.PREDICTION_TYPES`)
`forecasting_horizon`	integer	`None`	Steps to forecast; `None` uses the fitted horizon
`coverage_rates`	list of float	`[]`	Interval coverage rates; `[]` uses the forecaster's default. `interval` only
`groups`	list of string	`[]`	Panel groups to predict; `[]` predicts all groups

The default prediction_type is point when the forecaster supports it, otherwise class_proba, otherwise interval.

Behaviour¶

Each call works on a deep copy of the loaded forecaster. Changes made during a call are not visible to later calls.

When y is present, call observe(y, X_actual=..., groups=..., X_future=..., X_forecast=...).
Call the method for prediction_type with X_future, X_forecast, forecasting_horizon and groups:

`prediction_type`	Method	Extra argument
`point`	`predict`
`interval`	`predict_interval`	`coverage_rates`
`class_proba`	`predict_class_proba`

Empty coverage_rates and groups are passed as None.

Output¶

The polars DataFrame returned by the method, unchanged. With y, it holds one vintage, at the last observed time.

Errors¶

Condition	Exception
`model_input` is not a dict (for example a pandas DataFrame)	`TypeError`
A key is not in `INPUT_KEYS`	`ValueError`
A value is not a polars DataFrame	`TypeError`
`X_actual` without `y`	`ValueError`
`prediction_type` not supported by the forecaster	`ValueError`
Non-empty `coverage_rates` with a `prediction_type` other than `interval`	`ValueError`

Signature rules at save time¶

Signature passed to `save_model`	Result
None	Params-only signature with the four params
Without params or inputs	The four params are added; outputs are kept
With other params	Merged with the four params
With a param named like one of the four	`ValueError`; nothing is written
With an input schema	`ValueError`; nothing is written

Not supported¶

pandas input, and therefore mlflow models serve and Spark UDFs.