_base.py

"""Private base classes for tpcp.

These classes are in a separate module to avoid circular imports.
In basically all cases, you do not need them.
"""
from __future__ import annotations

import contextlib
import copy
import dataclasses
import inspect
import os
import sys
import warnings
from collections import defaultdict
from functools import wraps
from pathlib import Path
from types import MethodWrapperType
from typing import (
    TYPE_CHECKING,
    Annotated,
    Any,
    Callable,
    ClassVar,
    Generic,
    Optional,
    TypeVar,
    Union,
    cast,
)

import numpy as np
import pandas as pd
from sklearn.base import BaseEstimator
from typing_extensions import (
    Concatenate,
    Literal,
    ParamSpec,
    Self,
    get_args,
    get_origin,
)

from tpcp._parameters import _ParaTypes
from tpcp.exceptions import (
    MutableDefaultsError,
    PotentialUserErrorWarning,
    ValidationError,
)

try:
    import tensorflow as tf
except ImportError:
    tf = None

if TYPE_CHECKING:
    from collections.abc import Iterable

T = TypeVar("T")
P = ParamSpec("P")
BaseTpcpObjectT = TypeVar("BaseTpcpObjectT", bound="BaseTpcpObject")
_BaseTpcpObjectT = TypeVar("_BaseTpcpObjectT", bound="_BaseTpcpObject")


class _Nothing:
    """Sentinel class to indicate the lack of a value when ``None`` is ambiguous.

    ``_Nothing`` is a singleton. There is only ever one of it.

    This implementation is taken from the attrs package.
    """

    _singleton: Optional[_Nothing] = None

    def __new__(cls) -> _Nothing:
        if _Nothing._singleton is None:
            _Nothing._singleton = super().__new__(cls)
        return _Nothing._singleton

    def __repr__(self) -> str:
        return "NOTHING"

    def __bool__(self) -> Literal[False]:
        return False

    def __len__(self) -> Literal[0]:
        return 0  # __bool__ for Python 2


NOTHING = _Nothing()
"""
Sentinel to indicate the lack of a value when ``None`` is ambiguous.
"""


def _get_init_defaults(cls: type[_BaseTpcpObject]) -> dict[str, inspect.Parameter]:
    # fetch the constructor or the original constructor before deprecation wrapping if any
    init = cls.__init__
    if init is object.__init__ or isinstance(init, MethodWrapperType):
        # No explicit constructor to introspect
        return {}

    # introspect the constructor arguments to find the model parameters to represent
    init_signature = inspect.signature(init)
    # Consider the constructor parameters excluding 'self'
    defaults = {k: p for k, p in init_signature.parameters.items() if p.name != "self" and p.kind != p.VAR_KEYWORD}
    return defaults


def _replace_defaults_wrapper(
    cls: type[_BaseTpcpObjectT], old_init: Callable[Concatenate[_BaseTpcpObjectT, P], T]
) -> Callable[Concatenate[_BaseTpcpObjectT, P], T]:
    """Decorate an init to create new instances of mutable defaults.

    This should only be used in combination with `default` and will be applied as part of `__init_subclass`.
    Direct usage of this decorator should not be required.
    """
    # We get the params of the old inits here and not at runtime to avoid issues when the class is subclassed and
    # super().__init__ is called before all parameters of the child object are set.
    # This way, the param checks only concern the parameters of the current class.
    params = get_param_names(cls)

    @wraps(old_init)
    def new_init(self: _BaseTpcpObjectT, *args: P.args, **kwargs: P.kwargs) -> None:
        # call the old init.
        old_init(self, *args, **kwargs)

        # Check if any of the initial values has a "default parameter flag".
        # If yes we replace it with a clone (in case of a tpcp object) or a deepcopy in case of other objects.
        # This is handled by the factory `get_value` method.
        for p in params:
            if isinstance(val := getattr(self, p), BaseFactory):
                setattr(self, p, val.get_value())

    # This is just for introspection, in case we want to know if we have a modified init.
    new_init.__tpcp_wrapped__ = True

    return cast(Callable[Concatenate[_BaseTpcpObjectT, P], T], new_init)


def _retry_eval_with_missing_locals(
    expression: str,
    globalns: Optional[dict[str, Any]] = None,
    localns: Optional[dict[str, Any]] = None,
) -> Any:
    globalns = globalns or {}
    localns = localns or {}
    # We make a copy to not overwrite the input dict
    localns = {**localns}

    # That seems scary :D Let's see if this causes any issues
    # We use a value here instead of a "while True" to not get the program stuck in an endless loop.
    for _ in range(100):
        try:
            val = eval(expression, globalns, localns)  # noqa: PGH001
            break
        except NameError as e:
            missing = str(e).split("'")[1]
            localns[missing] = None
        except AttributeError as e:
            raise RuntimeError(
                "You ran into an edge case of the builtin type resolver. "
                "This happens if you use a nested type annotation that is only valid during runtime. "
                "This usually happens if you are using a `if TYPE_CHECKING:` guard for some of your imports to avoid "
                "circular dependencies.\n"
                "For most of these cases we have a built in workaround, but only if you provide the type directly and "
                "not in an attribute notation:\n"
                "\n"
                "This is not allowed:"
                "\n"
                ">>> if TYPE_CHECKING:\n"
                "...     import my\n"
                ">>> MyClass:\n"
                "...     para: my.custom_type\n"
                "\n\n"
                "This should work:\n"
                ">>> if TYPE_CHECKING:\n"
                "...     from my import custom_type\n"
                ">>> MyClass:\n"
                "...     para: custom_type\n"
                "\n\n"
                "And this as well:\n"
                ">>> if TYPE_CHECKING:\n"
                "...     import my\n"
                "...     custom_type_var = my.custom_type\n"
                ">>> MyClass:\n"
                "...     para: custom_type_var\n"
            ) from e
    else:
        raise RuntimeError(
            "Trying to resolve Parameter Type hints has resulted in an unexpected issue. "
            f"We were trying to evaluate the expression: {expression} . "
            "It seems like this expression contained forward references that could not be resolved. "
            "This should not happen!"
            "Please open an issue on GitHub with an code example that results in this issue. "
            "Sry for the inconvenience :)"
        )
    return val


def _custom_get_type_hints(cls: type[_BaseTpcpObject]) -> dict[str, Any]:
    """Extract type hints while avoiding issues with forward references.

    We automatically skip all douple-underscore methods.
    """
    hints = {}
    for base in reversed(cls.__mro__):
        base_globals = sys.modules[base.__module__].__dict__
        ann = base.__dict__.get("__annotations__", {})
        for name, value in ann.items():
            if name.startswith("__"):
                continue
            if value is None:
                value = type(None)  # noqa: PLW2901
            elif isinstance(value, str):
                # NOTE: This does not check if the str is a valid expression.
                #       This might not be an issue, but could lead to obscure error messages.
                value = _retry_eval_with_missing_locals(value, base_globals)  # noqa: PLW2901
            hints[name] = value
    return hints


def _extract_annotations(
    cls: type[_BaseTpcpObject], init_fields: dict[str, inspect.Parameter]
) -> dict[str, _ParaTypes]:
    cls_annotations = _custom_get_type_hints(cls)
    para_annotations = {}
    for k, v in cls_annotations.items():
        origin = get_origin(v)
        if origin is ClassVar:
            # If the parameter is a ClassVar, we go one level deeper and check, if its argument was annotated.
            v = get_args(v)[0]  # noqa: PLW2901
            origin = get_origin(v)
        if origin is Annotated:
            for annot in get_args(v)[1:]:
                if isinstance(annot, _ParaTypes):
                    para_annotations[k] = annot
                    break
        elif k in init_fields:
            para_annotations[k] = _ParaTypes.SIMPLE
    return para_annotations


def _validate_parameter(cls: type[_BaseTpcpObject]):
    # We extract the fields of the init
    fields = _get_init_defaults(cls)

    # Validation
    _has_dangerous_mutable_default(fields, cls)
    _has_invalid_name(fields, cls)

    if (
        dataclasses.is_dataclass(cls)
        or getattr(
            cls, "__attrs_attrs__", False
        )  # This checks if the class it an `attrs` class, without importing `attrs`
    ) and any(isinstance(field.default, BaseFactory) for field in fields.values()):
        # When you are already using dataclasses or attrs, you must use their factory methods, as we don't overwrite
        # the init to support CloneFactory anymore.
        raise ValidationError(
            "You are using the tpcp default factory (`cf`/`CloneFactory`) in combination with "
            "`dataclasses` or `attrs`. "
            "Use the `default_factory`/`factory` option of `dataclasses.field`/`attrs.field` instead."
        )


def _validate_all_parent_parameters_implemented(cls: type[_BaseTpcpObject]):
    """Validate that a class implements all parameters of its parents.

    We don't raise an error in this case, we just warn the user.
    """
    param_names = set(get_param_names(cls))

    for base in cls.__bases__:
        if not issubclass(base, _BaseTpcpObject):
            continue
        base_params = set(get_param_names(base))
        if not base_params.issubset(param_names):
            missing_params = base_params - param_names
            warnings.warn(
                f"The class {cls.__name__} does not implement all parameters of its parent {base.__name__}. "
                f"Missing parameters: {missing_params}\n"
                "This might not be a problem, but indicates bad design and you might run into actual issues with some "
                "of the validation magic `tpcp` does in the background. "
                "We would recommend to implement all parameters of your parents in a subclass.",
                stacklevel=2,
            )


def _get_tpcp_validated(cls_or_instance: Union[type[_BaseTpcpObject], _BaseTpcpObject]):
    cls = cls_or_instance if isinstance(cls_or_instance, type) else type(cls_or_instance)
    return cls.__dict__.get("__tpcp_validated_hidden__", False)


def _set_tpcp_validated(cls_or_instance: Union[type[_BaseTpcpObject], _BaseTpcpObject], value: bool):
    cls = cls_or_instance if isinstance(cls_or_instance, type) else type(cls_or_instance)
    cls.__tpcp_validated_hidden__ = value


class _BaseTpcpObject:
    # These two parameters should be initialized once per class.
    # This means, when we read them, we always check in the dictionary if they are defined.
    __field_annotations_cache__: dict[str, _ParaTypes]
    __tpcp_validated_hidden__: bool

    @property
    def __field_annotations__(self) -> dict[str, _ParaTypes]:
        """Get the field annotations that provide higher level information about the role of a parameter.

        We implement that as property to move the extraction of the annotations as far down in the initialization as
        possible, when we are sure the cls definition is properly finalized (i.e. dataclass decorators are run)
        """
        cls = type(self)
        # We check if we have a cache, and if it is our cache or the cache inherited by the parent class.
        # If it is from the parent class, we make sure, we create our own.
        if cache := cls.__dict__.get("__field_annotations_cache__", None):
            return cache
        # We need to create a new cache!
        fields = _get_init_defaults(cls)
        field_annotations = _extract_annotations(cls, fields)
        # We only validate the annotations here. If the user is not using the annotations, no need to bother them
        # with useless error messages
        _annotations_are_valid(field_annotations, fields, cls_name=cls.__name__)
        # We set the cache on the cls, so that only one instance is required to process it.
        cls.__field_annotations_cache__ = field_annotations
        return field_annotations

    def __init_subclass__(cls, **kwargs: Any):
        super().__init_subclass__(**kwargs)
        # If the class has no init or did inherit its init from the parent (i.e. it is not in its own dict), there is
        # nothing to do.
        if "__init__" not in cls.__dict__:
            return

        # If we have a custom init, we need to wrap it with a method that replaces our default values on init.
        setattr(cls, "__init__", _replace_defaults_wrapper(cls, cls.__init__))

    @classmethod
    def __custom_tpcp_validation__(cls):
        """Trigger custom validation.

        This method can be overwritten by the user to implement custom validation.
        The validation is triggered on the first call of `get_params.
        Within the method, either raise a `ValidationError` or privide warnings to the user.
        """
        return


class BaseTpcpObject(_BaseTpcpObject):
    """Baseclass for all tpcp objects."""

    _composite_params: ClassVar[tuple[str, ...]] = ()

    def get_params(self, deep: bool = True) -> dict[str, Any]:
        """Get parameters for this algorithm.

        Parameters
        ----------
        deep
            Only relevant if object contains nested algorithm objects.
            If this is the case and deep is True, the params of these nested objects are included in the output using a
            prefix like `nested_object_name__` (Note the two "_" at the end)

        Returns
        -------
        params
            Parameter names mapped to their values.

        """
        return _get_params(self, deep)

    def set_params(self, **params: Any) -> Self:
        """Set the parameters of this Algorithm.

        To set parameters of nested objects use `nested_object_name__para_name=`.
        """
        return _set_params(self, **params)

    def clone(self) -> Self:
        """Create a new instance of the class with all parameters copied over.

        This will create a new instance of the class itself and all nested objects
        """
        return clone(self, safe=True)

    def __repr_parameter__(self, name: str, value: Any) -> str:
        """Format a parameter for the __repr__ method.

        This method is used to format the parameters for the __repr__ method.
        It is called for each parameter and can be overwritten to customize the output.

        Parameters
        ----------
        name
            The name of the parameter
        value
            The value of the parameter

        Returns
        -------
        str
            The formatted string for the parameter

        """
        return f"{name}={value!r}"

    def __repr__(self) -> str:
        """Provide generic representation for the object based on all parameters."""
        class_name = type(self).__name__
        paras = self.get_params(deep=False)
        result = [class_name, "("]
        first = True
        for name, para in paras.items():
            if first:
                first = False
            else:
                result.append(", ")
            result.append(self.__repr_parameter__(name, para))
        return "".join(result) + ")"

    @classmethod
    def __clone_param__(cls, param_name: str, value: Any) -> Any:  # pylint: disable=unused-argument
        """Handle cloning of specific parameters of the object.

        This method exists only to allow object the handling of "strange" python datatypes that can not be deepcopied
        for some reason.
        This can happen when library authors overwrite deepcopy behaviour or if handling objects can not be easily
        deepcopied as there are implemented in Cython (or similar).

        In such cases you can overwrite how certain objects should be cloned using this method.
        Note, that you are on your own, when you do that.
        I.e. we will not check that you do it correctly and that you ensure that the new object is really independent of
        the old or any other things you might run into.

        If you are using this method, we recommend to at least check that the cloned object is identical to the
        original based on our implemented hashing method.

        >>> from tpcp._hash import custom_hash
        >>> assert custom_hash(my_obj) == custom_hash(my_obj.clone())

        If this doesn't pass, many other features in tpcp will not work as well.
        """
        return clone(value, safe=False)


def _get_deep_params(obj, parent_key) -> dict[str, Any]:
    # This is a little magic, that also gets the parameters of nested sklearn classifiers.
    if hasattr(obj, "get_params"):
        deep_items = obj.get_params(deep=True).items()
        return {parent_key + "__" + k: val for k, val in deep_items}
    return {}


def _assert_is_allowed_composite_value(val, parent_key: str, iteration: int):
    if (not isinstance(val, tuple)) or (len(val) != 2) or (not isinstance(val[0], str)):
        raise ValidationError(
            f"The provided parameters for the composite field {parent_key} does not seem to be the "
            "right type. "
            "It should be a sequence of `(name, value)` tuples, but the obj at position "
            f"{iteration} in the sequence was not a tuple but:\n`{val}`"
        )


def _recursive_validate(cls: type[_BaseTpcpObject]):
    if not _get_tpcp_validated(cls):
        _validate_parameter(cls)
        _validate_all_parent_parameters_implemented(cls)
        cls.__custom_tpcp_validation__()

        _set_tpcp_validated(cls, True)

        for base in cls.__bases__:
            if issubclass(base, _BaseTpcpObject):
                _recursive_validate(base)


def _get_params(instance: _BaseTpcpObject, deep: bool = True) -> dict[str, Any]:
    # At some point, we want to validate that the user defined the class correctly.
    # To allow for all strange modifications of the class, we run the validation, when we actually need the parameters.
    # This usually required a call to `get_params`.
    # Hence, we run the validation here.
    # To avoid running this validation on every get_params call, we use `__tpcp_validated__` to make sure it runs only
    # once per class.
    # We actually run this validation recusively for all base classes, in case `get_params` is never called on the base
    # directly.
    _recursive_validate(type(instance))

    valid_fields = get_param_names(type(instance))
    comp_fields = getattr(instance, "_composite_params", ())
    out: dict[str, Any] = {}
    for key in valid_fields:
        value = getattr(instance, key)
        if deep:
            if value is not None and comp_fields and key in comp_fields:
                # Here we handle composite field (i.e. fields that contain sequences of (name, tpcp_obj_instance)
                # tuples.
                # We basically flatten the list by storing the instance under the name `{key}__{name}` and then add
                # all the nested parameters
                for i, v in enumerate(value):
                    _assert_is_allowed_composite_value(v, key, i)
                    name, obj = v
                    nested_key = f"{key}__{name}"
                    out[nested_key] = obj
                    if deep:
                        out.update(**_get_deep_params(obj, nested_key))
            else:
                out.update(**_get_deep_params(value, key))
        out[key] = value
    return out


def _set_comp_field(instance, field_name, params):
    # We first partition our field names to know to which index they belong
    comp_params: defaultdict[str, Any] = defaultdict(dict)
    for key, value in params.items():
        key, delim, sub_key = key.partition("__")  # noqa: PLW2901
        if delim:
            comp_params[key][sub_key] = value
        else:
            comp_params[key]["*"] = value

    # Then we iterate over existing values in the compound field and recreate it
    new_list = []
    comp_list = list(getattr(instance, field_name))
    for key, old_value in comp_list:
        nested_values = comp_params.pop(key, {})
        new_value = old_value if "*" not in nested_values else nested_values.pop("*")
        if nested_values:
            new_value = new_value.set_params(**nested_values)
        new_list.append((key, new_value))
    if comp_params:
        # Some values are left over -> aka they are invalid
        raise ValueError(
            f"You are trying to set values for a compound field {field_name} with the identifiers "
            f"{list(comp_params.keys())}. "
            "These identifiers could not be found within the existing list of identifiers for this "
            f"field ({[c[0] for c in comp_list]}. "
            "We don't support setting params on not existent value in a compound field. "
            "Recreate the entire field if you want to add or delete entries from a compound field."
        )

    setattr(instance, field_name, new_list)


def _set_params(instance: BaseTpcpObjectT, **params: Any) -> BaseTpcpObjectT:
    """Set the parameters of an instance.

    To set parameters of nested objects use `nested_object_name__para_name=`.
    """
    # Basically copied from sklearn
    if not params:
        return instance
    valid_params = instance.get_params(deep=True)
    comp_fields = getattr(instance, "_composite_params", ())

    nested_params: defaultdict[str, Any] = defaultdict(dict)  # grouped by prefix
    for key, value in params.items():
        key, delim, sub_key = key.partition("__")  # noqa: PLW2901
        if key not in valid_params:
            raise ValueError(f"`{key}` is not a valid parameter name for {type(instance).__name__}.")

        if delim:
            nested_params[key][sub_key] = value
        else:
            setattr(instance, key, value)
            valid_params[key] = value

    for key, sub_params in nested_params.items():
        if key in comp_fields:
            _set_comp_field(instance, key, sub_params)
        else:
            valid_params[key].set_params(**sub_params)
    return instance


def get_param_names(obj: Union[type[_BaseTpcpObject], _BaseTpcpObject]) -> list[str]:
    """Get parameter names for the object.

    The parameters of an algorithm/pipeline are defined based on its `__init__` method.
    All parameters of this method are considered parameters of the algorithm.

    Notes
    -----
    Adopted based on :meth:`sklearn.base.BaseEstimator._get_param_names`.

    Returns
    -------
    param_names
        List of parameter names of the algorithm

    """
    cls = obj if isinstance(obj, type) else type(obj)
    parameters = list(_get_init_defaults(cls).values())
    for p in parameters:
        if p.kind == p.VAR_POSITIONAL:
            raise RuntimeError(
                "tpcp algorithms and pipelines should always specify their parameters in the signature of their "
                f"__init__ (no varargs). {cls} doesn't follow this convention."
            )
    # Extract and sort argument names excluding 'self'
    return sorted([p.name for p in parameters])


def _get_annotated_fields_of_type(
    instance_or_cls: BaseTpcpObject, field_type: Union[_ParaTypes, Iterable[_ParaTypes]]
) -> list[str]:
    if isinstance(field_type, _ParaTypes):
        field_type = [field_type]
    return [k for k, v in instance_or_cls.__field_annotations__.items() if v in field_type]


def _has_dangerous_mutable_default(fields: dict[str, inspect.Parameter], cls: type[_BaseTpcpObject]) -> None:
    mutable_defaults = []

    for name, field in fields.items():
        if _is_dangerous_mutable(field):
            # We do not raise an error right here, but wait until we checked everything to provide a more helpful
            # error message.
            mutable_defaults.append(name)

    if len(mutable_defaults) > 0:
        raise MutableDefaultsError(
            f"The class {cls.__name__} contains mutable objects as default values ({mutable_defaults}). "
            "This can lead to unexpected and unpleasant issues! "
            "To solve this the default value should be generated by a factory that produces a new value for each "
            "instance. "
            "If you simply want a new version of the provided default, wrap the value into `tpcp.CloneFactory`"
            "This will enforce a clone/copy of the respective input, whenever a new instance of your Algorithm "
            "or Pipeline is created. "
            "\n"
            "Note, that we do not check for all cases of mutable objects. "
            f"At the moment, we check only for {_get_dangerous_mutable_types()}. "
            "To learn more about this topic, check TODO: LINK."
        )


def _annotations_are_valid(
    field_annotations: dict[str, _ParaTypes],
    fields: dict[str, inspect.Parameter],
    cls_name: str,
) -> None:
    for k, v in field_annotations.items():
        if "__" in k:
            if v is _ParaTypes.SIMPLE:
                warnings.warn(
                    "Annotating a nested parameter (parameter like `nested_object__nest_para` as a simple "
                    "Parameter has no effect and the entire line should be removed.",
                    PotentialUserErrorWarning,
                    stacklevel=2,
                )
        elif k not in fields:
            raise ValueError(
                f"The field '{k}' of {cls_name} was annotated as a `tpcp` "
                "(Hyper/Pure/Normal/Optimizable)-Parameter, but is not a parameter listed in the init! "
                "Add the parameter to the init if it is an actual parameter of your algorithm, or remove the "
                "annotation."
            )


def _has_invalid_name(fields: dict[str, inspect.Parameter], cls: type[_BaseTpcpObject]) -> None:
    invalid_names = [f for f in fields if "__" in f]
    if len(invalid_names) > 0:
        raise ValidationError(
            f"The parameters {invalid_names} of {cls.__name__} have a double-underscore in their name. "
            "This is not allowed, as it interferes with the nested naming conventions in tpcp.\n\n"
            "If you are seeing this while using `dataclasses`, and trying to annotate nested parameters, make sure to "
            "exclude from the init using the explicit `field` syntax or by wrapping it with `ClassVar`:\n\n"
            ">>> @dataclasses.dataclass\n"
            "... class MyClass(BaseTpcpObject):\n"
            "...    # first option\n"
            "...    nested__parameter: OptiPara[int] = dataclasses.field(init=False)\n"
            "...    # second option\n"
            "...    other_nested__parameter: ClassVar[OptiPara[int]]\n\n"
        )
    invalid_names = [f for f in fields if f.endswith("_")]
    if len(invalid_names) > 0:
        raise ValidationError(
            f"The parameters {invalid_names} of {cls.__name__} have a trailing underscore in their name. "
            "This is not allowed, as it interferes naming convention of result objects in tpcp. "
            "Only result values are allowed to have a trailing underscore.\n\n"
            "If you are seeing this while using `dataclasses` or `attrs` and trying to annotate types of result "
            "objects, make sure to exclude them from the init using the explicit `field` syntax:\n\n"
            ">>> @dataclasses.dataclass\n"
            "... class MyClass(BaseTpcpObject):\n"
            "...    result_: int = dataclasses.field(init=False)\n"
        )


def _get_dangerous_mutable_types() -> tuple[type, ...]:
    # TODO: Update this list or even make it a white list?
    return _BaseTpcpObject, list, dict, np.ndarray, pd.DataFrame, BaseEstimator


def _is_dangerous_mutable(field: inspect.Parameter) -> bool:
    """Check if a parameter is one of the mutable objects "considered" dangerous."""
    if field.default is inspect.Parameter.empty or isinstance(field.default, BaseFactory):
        return False
    if isinstance(field.default, _get_dangerous_mutable_types()):
        return True
    return False


def _is_builtin_class_instance(obj: Any) -> bool:
    return type(obj).__module__ == "builtins"


def clone(algorithm: T, *, safe: bool = False) -> T:
    """Construct a new algorithm object with the same parameters.

    This is a modified version from sklearn and the original was published under a BSD-3 license and the original file
    can be found here: https://github.com/scikit-learn/scikit-learn/blob/0d378913b/sklearn/base.py#L31

    The method creates a copy of tpcp algorithms and pipelines, without any results attached.
    I.e. it is equivalent to creating a new instance with the same parameter.
    For all objects that are not tpcp objects (or lists of them), a deepcopy is created.

    .. warning :: This function will not clone sklearn models as expected!
                  `sklearn.clone` will remove the trained model from sklearn object by creating a new instance.
                  This clone method, will deepcopy sklearn models.
                  This means fitted models will be copied and are still available afterwards.
                  For more information have a look at the documentation about "inputs and results" in tpcp.
                  (see :ref:`general concepts <general_concepts>` and :ref:`sklearn differences <sklearn_differences>`)

    Parameters
    ----------
    algorithm : {list, tuple, set} of algorithm instance or a single algorithm instance
        The algorithm or group of algorithms to be cloned.
    safe : bool, default=False
        If safe is False, clone will fall back to a deep copy on objects that are not algorithms.

    """
    if algorithm is NOTHING:
        return algorithm
    # Handle named tuple
    if isinstance(algorithm, tuple) and hasattr(algorithm, "_asdict") and hasattr(algorithm, "_fields"):
        return type(algorithm)(*(clone(a, safe=safe) for a in algorithm))
    if isinstance(algorithm, dict):
        return {k: clone(v, safe=safe) for k, v in algorithm.items()}
    if isinstance(algorithm, (list, tuple, set, frozenset)):
        return type(algorithm)([clone(a, safe=safe) for a in algorithm])

    # Compared to sklearn, we check specifically for BaseTpcpObject and not just if `get_params` is defined on the
    # object.
    # Due to the way algorithms/pipelines in tpcp work, they need to inherit from BaseTpcpObject.
    # Therefore, we check explicitly for that, as we do not want to accidentally treat a sklearn algo (or similar) as
    # algorithm
    if not isinstance(algorithm, BaseTpcpObject):
        if not safe:
            # For some reason, some libraries print stuff to stdout when cloning.
            with Path(os.devnull).open("w") as devnull, contextlib.redirect_stdout(devnull):
                return copy.deepcopy(algorithm)
        raise TypeError(
            f"Cannot clone object '{algorithm!r}' (type {type(algorithm)}): "
            "it does not seem to be a compatible algorithm/pipline class or general `tpcp` object as it does not "
            "inherit from `BaseTpcpObject` or `Algorithm` or `Pipeline`."
        )

    klass = algorithm.__class__
    new_object_params = algorithm.get_params(deep=False)
    for name, param in new_object_params.items():
        # We defer the actual cloning of the parameters to the class itself to allow to handle stupid edge cases on
        # user side without modifying tpcp itself.
        new_object_params[name] = klass.__clone_param__(name, param)
    new_object = klass(**new_object_params)
    params_set = new_object.get_params(deep=False)

    # quick sanity check of the parameters of the clone
    for name in new_object_params:
        param1 = new_object_params[name]
        param2 = params_set[name]
        if param1 is not param2:
            raise RuntimeError(
                f"Cannot clone object {algorithm}, as the constructor either does not set or modifies parameter {name}"
            )
    return new_object


class BaseFactory:
    """Baseclass for factories to circumvent mutable defaults."""

    def get_value(self) -> Any:
        """Provide the value generated by the factory.

        This method will be called every time a new instance of a class is created.
        """
        raise NotImplementedError()

    def __call__(self) -> Any:
        return self.get_value()


class CloneFactory(BaseFactory, Generic[T]):
    """Init factory that creates a clone of the provided default values on instance initialisation.

    This can be used to make sure that each instance get their own version own copy of a mutable default of a class
    init.
    Under the hood this uses :func:`~tpcp.clone`.
    """

    def __init__(self, default_value: T) -> None:
        self.default_value = default_value

    def get_value(self) -> T:
        """Clone the default value for each instance."""
        return clone(self.default_value)

    def __repr__(self) -> str:
        """Print the representation for the factory."""
        return f"cf({self.default_value!r})"


# The typing here is obviously wrong and a hack to make object that are wrapped by clone factory seem to be just of
# the same type.
# Ideally clone factory would be some generic type of proxy, but it isn't.
# Therefore, we fake the return type, so that external type checking in user code works without any errors.
# As long as `cf` is used for its intended purpose that should not matter much.
def cf(default_value: T) -> T:  # pylint: disable=invalid-name
    """Wrap mutable default value with the `CloneFactory`.

    This is basically an alias for :class:`~tpcp.CloneFactory`
    """
    return CloneFactory(default_value)  # type: ignore  # noqa: PGH003


class _Default:
    """Wrapper for default values.

    This can be used to check if an optional function argument was explicitly provided by the user or not.
    See :func:`~tpcp.validate.validate` for  a usage example.
    """

    def __init__(self, value):
        self.value = value

    def get_value(self):
        return self.value

    def __repr__(self):
        """Print the representation of the value."""
        return repr(self.value)