Source code for qsarkit.base._compat
"""Bridging the scikit-learn versions the package supports.
``requires-python`` is ``>=3.9`` and the dependency floor is
``scikit-learn>=1.2``, which spans several years of releases. Parameters come
and go across that range: ``MLPRegressor`` gained ``loss`` and ``MDS`` gained
``init`` in versions that require Python 3.10 or newer, so code passing them
unconditionally raises ``TypeError`` on exactly the interpreter the package
claims to support.
Forwarding such a parameter only when the installed estimator accepts it keeps
one code path working across the whole range. Silently dropping a value the
caller asked for would be worse than failing, so a non-default value on a
version that cannot honour it raises with a message naming the version.
"""
from __future__ import annotations
import inspect
from functools import lru_cache
from typing import Any, Dict, Mapping
__all__ = ["accepts_parameter", "forward_supported"]
[docs]
@lru_cache(maxsize=None)
def accepts_parameter(cls: type, name: str) -> bool:
"""Whether ``cls.__init__`` takes a parameter called ``name``.
Cached, because it is consulted on every construction and the answer
cannot change within a process.
Parameters
----------
cls : type
name : str
Returns
-------
bool
Examples
--------
>>> from sklearn.ensemble import RandomForestRegressor
>>> from qsarkit.base import accepts_parameter
>>> accepts_parameter(RandomForestRegressor, "n_estimators")
True
>>> accepts_parameter(RandomForestRegressor, "not_a_parameter")
False
"""
try:
# The class itself, not cls.__init__: this gives the constructor's
# signature without `self`, and does not depend on which class in the
# hierarchy supplied __init__.
signature = inspect.signature(cls)
except (TypeError, ValueError): # pragma: no cover - C-level constructor
return False
parameters = signature.parameters
if name in parameters:
return True
# A constructor taking **kwargs accepts anything.
return any(p.kind is inspect.Parameter.VAR_KEYWORD for p in parameters.values())
[docs]
def forward_supported(
cls: type, candidates: Mapping[str, Any], defaults: Mapping[str, Any]
) -> Dict[str, Any]:
"""Keep the entries ``cls`` accepts; refuse to drop a non-default silently.
Parameters
----------
cls : type
The estimator being constructed.
candidates : mapping
Parameters to forward if supported.
defaults : mapping
The value each candidate has by default. A candidate equal to its
default is dropped quietly when unsupported, because the caller did
not ask for it; anything else raises.
Returns
-------
dict
The subset of ``candidates`` that ``cls`` accepts.
Raises
------
TypeError
If a candidate differs from its default and ``cls`` cannot accept it.
Examples
--------
>>> from sklearn.ensemble import RandomForestRegressor
>>> from qsarkit.base import forward_supported
>>> forward_supported(
... RandomForestRegressor,
... {"n_estimators": 50, "nonexistent": "default"},
... {"n_estimators": 100, "nonexistent": "default"},
... )
{'n_estimators': 50}
"""
import sklearn
forwarded: Dict[str, Any] = {}
for name, value in candidates.items():
if accepts_parameter(cls, name):
forwarded[name] = value
elif value != defaults.get(name):
raise TypeError(
f"{cls.__name__} in scikit-learn {sklearn.__version__} has no "
f"{name!r} parameter, so {name}={value!r} cannot be honoured. "
f"Upgrade scikit-learn, or leave {name} at its default."
)
return forwarded