Saved Model Format¶

Description of a model directory written by save_model or log_model, and of the checks load_model and check_compatibility apply to it. Format version 1.0 (yohou_mlflow.FORMAT_VERSION).

Directory layout¶

MLmodel: flavors.yohou¶

Example:

flavors:
  yohou:
    components:
      forecaster: forecaster.skops
    default_prediction_type: point
    forecaster_class: yohou.interval.split_conformal.SplitConformalForecaster
    forecaster_type:
    - interval
    - point
    format_version: '1.0'
    recorded_types:
    - datetime.datetime
    - polars.dataframe.frame.DataFrame
    - polars.datatypes.classes.Float64
    - yohou.interval.split_conformal.SplitConformalForecaster
    - yohou.metrics.conformity.Residual
    - yohou.point.reduction.PointReductionForecaster
    versions:
      polars: 1.44.2
      scikit-learn: 1.9.1
      skops: 0.15.0
      yohou: 0.1.0a13
      yohou-mlflow: 0.1.0

MLmodel: flavors.python_function¶

The model signature declares params and no inputs. See pyfunc contract.

Format version¶

A component file name that is empty, contains a path separator, or is . or .., raises YohouMlflowError.

Trust policy¶

A type in forecaster.skops is trusted when any of the following holds:

The prefixes and exact names are exported as yohou_mlflow.TRUSTED_TYPE_PREFIXES and yohou_mlflow.TRUSTED_TYPES.

polars frames are stored in polars' binary format and rebuilt by polars.DataFrame.deserialize, which skops does not inspect.

Version rules¶

The rules are exported as yohou_mlflow.VERSION_RULES.

extra_version_rules adds rules for other packages, as a mapping of distribution name to "exact" or "major.minor", on save_model, log_model, load_model and check_compatibility, and as a model_config key of mlflow.pyfunc.load_model. At save, a declared package is recorded in versions only when forecaster.skops holds a type from one of its top-level modules. At load, a recorded package is compared only under a rule the loading call declares. A rule for a package in VERSION_RULES, or a rule other than "exact" and "major.minor", raises ValueError.

With strict=True (default), a mismatch raises VersionMismatchError. With strict=False, it emits VersionMismatchWarning with the same text and the load proceeds.

Load order¶

load_model applies these steps in order. Each failing step stops the load before any object is constructed from forecaster.skops:

1. Read flavors.yohou; check format_version.
2. Compare versions under the version rules.
3. Read the types in forecaster.skops with skops.io.get_untrusted_types and check them against the trust policy; raise UntrustedTypesError for any outside it.
4. Load with skops.io.load, trusting exactly those types.

check_compatibility applies steps 1 to 3 and returns a CompatibilityReport instead of raising.

Neither function puts any directory from the model on Python's import path, and a code key in the MLmodel file is ignored.

mlflow.pyfunc.load_model is outside this guarantee: MLflow imports the loader_module and any code directory named in flavors.python_function before calling this package.

Save checks¶

save_model applies the first check before anything is written, and the others after writing forecaster.skops, removing the directory if one fails:

When the original's default predict method raises without inputs, the comparison is skipped and a UserWarning is emitted.

Default pip requirements¶

yohou_mlflow, yohou, scikit-learn, polars and skops, each pinned to the installed version (get_default_pip_requirements). pip_requirements replaces them; extra_pip_requirements adds to them.

Time zones¶

A forecaster fitted on time-zone-aware data holds its time zone in observed_time_ as a zoneinfo.ZoneInfo, including fixed offsets such as +02:00, which polars returns as ZoneInfo too. It is saved and restored. This requires skops 0.16.0 or later, the package's minimum.

Errors¶