from abc import ABC from abc import abstractmethod from collections.abc import Iterable import functools import math from inspect import signature from numbers import Integral from numbers import Real import operator import re import warnings import numpy as np from scipy.sparse import issparse from scipy.sparse import csr_matrix from .validation import _is_arraylike_not_scalar class InvalidParameterError(ValueError, TypeError): """Custom exception to be raised when the parameter of a class/method/function does not have a valid type or value. """ # Inherits from ValueError and TypeError to keep backward compatibility. def validate_parameter_constraints(parameter_constraints, params, caller_name): """Validate types and values of given parameters. Parameters ---------- parameter_constraints : dict or {"no_validation"} If "no_validation", validation is skipped for this parameter. If a dict, it must be a dictionary `param_name: list of constraints`. A parameter is valid if it satisfies one of the constraints from the list. Constraints can be: - an Interval object, representing a continuous or discrete range of numbers - the string "array-like" - the string "sparse matrix" - the string "random_state" - callable - None, meaning that None is a valid value for the parameter - any type, meaning that any instance of this type is valid - an Options object, representing a set of elements of a given type - a StrOptions object, representing a set of strings - the string "boolean" - the string "verbose" - the string "cv_object" - the string "missing_values" - a HasMethods object, representing method(s) an object must have - a Hidden object, representing a constraint not meant to be exposed to the user params : dict A dictionary `param_name: param_value`. The parameters to validate against the constraints. caller_name : str The name of the estimator or function or method that called this function. """ for param_name, param_val in params.items(): # We allow parameters to not have a constraint so that third party estimators # can inherit from sklearn estimators without having to necessarily use the # validation tools. if param_name not in parameter_constraints: continue constraints = parameter_constraints[param_name] if constraints == "no_validation": continue constraints = [make_constraint(constraint) for constraint in constraints] for constraint in constraints: if constraint.is_satisfied_by(param_val): # this constraint is satisfied, no need to check further. break else: # No constraint is satisfied, raise with an informative message. # Ignore constraints that we don't want to expose in the error message, # i.e. options that are for internal purpose or not officially supported. constraints = [ constraint for constraint in constraints if not constraint.hidden ] if len(constraints) == 1: constraints_str = f"{constraints[0]}" else: constraints_str = ( f"{', '.join([str(c) for c in constraints[:-1]])} or" f" {constraints[-1]}" ) raise InvalidParameterError( f"The {param_name!r} parameter of {caller_name} must be" f" {constraints_str}. Got {param_val!r} instead." ) def make_constraint(constraint): """Convert the constraint into the appropriate Constraint object. Parameters ---------- constraint : object The constraint to convert. Returns ------- constraint : instance of _Constraint The converted constraint. """ if isinstance(constraint, str) and constraint == "array-like": return _ArrayLikes() if isinstance(constraint, str) and constraint == "sparse matrix": return _SparseMatrices() if isinstance(constraint, str) and constraint == "random_state": return _RandomStates() if constraint is callable: return _Callables() if constraint is None: return _NoneConstraint() if isinstance(constraint, type): return _InstancesOf(constraint) if isinstance(constraint, (Interval, StrOptions, Options, HasMethods)): return constraint if isinstance(constraint, str) and constraint == "boolean": return _Booleans() if isinstance(constraint, str) and constraint == "verbose": return _VerboseHelper() if isinstance(constraint, str) and constraint == "missing_values": return _MissingValues() if isinstance(constraint, str) and constraint == "cv_object": return _CVObjects() if isinstance(constraint, Hidden): constraint = make_constraint(constraint.constraint) constraint.hidden = True return constraint raise ValueError(f"Unknown constraint type: {constraint}") def validate_params(parameter_constraints): """Decorator to validate types and values of functions and methods. Parameters ---------- parameter_constraints : dict A dictionary `param_name: list of constraints`. See the docstring of `validate_parameter_constraints` for a description of the accepted constraints. Note that the *args and **kwargs parameters are not validated and must not be present in the parameter_constraints dictionary. Returns ------- decorated_function : function or method The decorated function. """ def decorator(func): # The dict of parameter constraints is set as an attribute of the function # to make it possible to dynamically introspect the constraints for # automatic testing. setattr(func, "_skl_parameter_constraints", parameter_constraints) @functools.wraps(func) def wrapper(*args, **kwargs): func_sig = signature(func) # Map *args/**kwargs to the function signature params = func_sig.bind(*args, **kwargs) params.apply_defaults() # ignore self/cls and positional/keyword markers to_ignore = [ p.name for p in func_sig.parameters.values() if p.kind in (p.VAR_POSITIONAL, p.VAR_KEYWORD) ] to_ignore += ["self", "cls"] params = {k: v for k, v in params.arguments.items() if k not in to_ignore} validate_parameter_constraints( parameter_constraints, params, caller_name=func.__qualname__ ) try: return func(*args, **kwargs) except InvalidParameterError as e: # When the function is just a wrapper around an estimator, we allow # the function to delegate validation to the estimator, but we replace # the name of the estimator by the name of the function in the error # message to avoid confusion. msg = re.sub( r"parameter of \w+ must be", f"parameter of {func.__qualname__} must be", str(e), ) raise InvalidParameterError(msg) from e return wrapper return decorator def _type_name(t): """Convert type into human readable string.""" module = t.__module__ qualname = t.__qualname__ if module == "builtins": return qualname elif t == Real: return "float" elif t == Integral: return "int" return f"{module}.{qualname}" class _Constraint(ABC): """Base class for the constraint objects.""" def __init__(self): self.hidden = False @abstractmethod def is_satisfied_by(self, val): """Whether or not a value satisfies the constraint. Parameters ---------- val : object The value to check. Returns ------- is_satisfied : bool Whether or not the constraint is satisfied by this value. """ @abstractmethod def __str__(self): """A human readable representational string of the constraint.""" class _InstancesOf(_Constraint): """Constraint representing instances of a given type. Parameters ---------- type : type The valid type. """ def __init__(self, type): super().__init__() self.type = type def is_satisfied_by(self, val): return isinstance(val, self.type) def __str__(self): return f"an instance of {_type_name(self.type)!r}" class _NoneConstraint(_Constraint): """Constraint representing the None singleton.""" def is_satisfied_by(self, val): return val is None def __str__(self): return "None" class _NanConstraint(_Constraint): """Constraint representing the indicator `np.nan`.""" def is_satisfied_by(self, val): return isinstance(val, Real) and math.isnan(val) def __str__(self): return "numpy.nan" class _PandasNAConstraint(_Constraint): """Constraint representing the indicator `pd.NA`.""" def is_satisfied_by(self, val): try: import pandas as pd return isinstance(val, type(pd.NA)) and pd.isna(val) except ImportError: return False def __str__(self): return "pandas.NA" class Options(_Constraint): """Constraint representing a finite set of instances of a given type. Parameters ---------- type : type options : set The set of valid scalars. deprecated : set or None, default=None A subset of the `options` to mark as deprecated in the string representation of the constraint. """ def __init__(self, type, options, *, deprecated=None): super().__init__() self.type = type self.options = options self.deprecated = deprecated or set() if self.deprecated - self.options: raise ValueError("The deprecated options must be a subset of the options.") def is_satisfied_by(self, val): return isinstance(val, self.type) and val in self.options def _mark_if_deprecated(self, option): """Add a deprecated mark to an option if needed.""" option_str = f"{option!r}" if option in self.deprecated: option_str = f"{option_str} (deprecated)" return option_str def __str__(self): options_str = ( f"{', '.join([self._mark_if_deprecated(o) for o in self.options])}" ) return f"a {_type_name(self.type)} among {{{options_str}}}" class StrOptions(Options): """Constraint representing a finite set of strings. Parameters ---------- options : set of str The set of valid strings. deprecated : set of str or None, default=None A subset of the `options` to mark as deprecated in the string representation of the constraint. """ def __init__(self, options, *, deprecated=None): super().__init__(type=str, options=options, deprecated=deprecated) class Interval(_Constraint): """Constraint representing a typed interval. Parameters ---------- type : {numbers.Integral, numbers.Real, "real_not_int"} The set of numbers in which to set the interval. If "real_not_int", only reals that don't have the integer type are allowed. For example 1.0 is allowed but 1 is not. left : float or int or None The left bound of the interval. None means left bound is -∞. right : float, int or None The right bound of the interval. None means right bound is +∞. closed : {"left", "right", "both", "neither"} Whether the interval is open or closed. Possible choices are: - `"left"`: the interval is closed on the left and open on the right. It is equivalent to the interval `[ left, right )`. - `"right"`: the interval is closed on the right and open on the left. It is equivalent to the interval `( left, right ]`. - `"both"`: the interval is closed. It is equivalent to the interval `[ left, right ]`. - `"neither"`: the interval is open. It is equivalent to the interval `( left, right )`. Notes ----- Setting a bound to `None` and setting the interval closed is valid. For instance, strictly speaking, `Interval(Real, 0, None, closed="both")` corresponds to `[0, +∞) U {+∞}`. """ def __init__(self, type, left, right, *, closed): super().__init__() self.type = type self.left = left self.right = right self.closed = closed self._check_params() def _check_params(self): if self.type not in (Integral, Real, "real_not_int"): raise ValueError( "type must be either numbers.Integral, numbers.Real or 'real_not_int'." f" Got {self.type} instead." ) if self.closed not in ("left", "right", "both", "neither"): raise ValueError( "closed must be either 'left', 'right', 'both' or 'neither'. " f"Got {self.closed} instead." ) if self.type is Integral: suffix = "for an interval over the integers." if self.left is not None and not isinstance(self.left, Integral): raise TypeError(f"Expecting left to be an int {suffix}") if self.right is not None and not isinstance(self.right, Integral): raise TypeError(f"Expecting right to be an int {suffix}") if self.left is None and self.closed in ("left", "both"): raise ValueError( f"left can't be None when closed == {self.closed} {suffix}" ) if self.right is None and self.closed in ("right", "both"): raise ValueError( f"right can't be None when closed == {self.closed} {suffix}" ) else: if self.left is not None and not isinstance(self.left, Real): raise TypeError("Expecting left to be a real number.") if self.right is not None and not isinstance(self.right, Real): raise TypeError("Expecting right to be a real number.") if self.right is not None and self.left is not None and self.right <= self.left: raise ValueError( f"right can't be less than left. Got left={self.left} and " f"right={self.right}" ) def __contains__(self, val): if np.isnan(val): return False left_cmp = operator.lt if self.closed in ("left", "both") else operator.le right_cmp = operator.gt if self.closed in ("right", "both") else operator.ge left = -np.inf if self.left is None else self.left right = np.inf if self.right is None else self.right if left_cmp(val, left): return False if right_cmp(val, right): return False return True def _has_valid_type(self, val): if self.type == "real_not_int": return isinstance(val, Real) and not isinstance(val, Integral) return isinstance(val, self.type) def is_satisfied_by(self, val): if not self._has_valid_type(val): return False return val in self def __str__(self): type_str = "an int" if self.type is Integral else "a float" left_bracket = "[" if self.closed in ("left", "both") else "(" left_bound = "-inf" if self.left is None else self.left right_bound = "inf" if self.right is None else self.right right_bracket = "]" if self.closed in ("right", "both") else ")" return ( f"{type_str} in the range " f"{left_bracket}{left_bound}, {right_bound}{right_bracket}" ) class _ArrayLikes(_Constraint): """Constraint representing array-likes""" def is_satisfied_by(self, val): return _is_arraylike_not_scalar(val) def __str__(self): return "an array-like" class _SparseMatrices(_Constraint): """Constraint representing sparse matrices.""" def is_satisfied_by(self, val): return issparse(val) def __str__(self): return "a sparse matrix" class _Callables(_Constraint): """Constraint representing callables.""" def is_satisfied_by(self, val): return callable(val) def __str__(self): return "a callable" class _RandomStates(_Constraint): """Constraint representing random states. Convenience class for [Interval(Integral, 0, 2**32 - 1, closed="both"), np.random.RandomState, None] """ def __init__(self): super().__init__() self._constraints = [ Interval(Integral, 0, 2**32 - 1, closed="both"), _InstancesOf(np.random.RandomState), _NoneConstraint(), ] def is_satisfied_by(self, val): return any(c.is_satisfied_by(val) for c in self._constraints) def __str__(self): return ( f"{', '.join([str(c) for c in self._constraints[:-1]])} or" f" {self._constraints[-1]}" ) class _Booleans(_Constraint): """Constraint representing boolean likes. Convenience class for [bool, np.bool_, Integral (deprecated)] """ def __init__(self): super().__init__() self._constraints = [ _InstancesOf(bool), _InstancesOf(np.bool_), _InstancesOf(Integral), ] def is_satisfied_by(self, val): # TODO(1.4) remove support for Integral. if isinstance(val, Integral) and not isinstance(val, bool): warnings.warn( "Passing an int for a boolean parameter is deprecated in version 1.2 " "and won't be supported anymore in version 1.4.", FutureWarning, ) return any(c.is_satisfied_by(val) for c in self._constraints) def __str__(self): return ( f"{', '.join([str(c) for c in self._constraints[:-1]])} or" f" {self._constraints[-1]}" ) class _VerboseHelper(_Constraint): """Helper constraint for the verbose parameter. Convenience class for [Interval(Integral, 0, None, closed="left"), bool, numpy.bool_] """ def __init__(self): super().__init__() self._constraints = [ Interval(Integral, 0, None, closed="left"), _InstancesOf(bool), _InstancesOf(np.bool_), ] def is_satisfied_by(self, val): return any(c.is_satisfied_by(val) for c in self._constraints) def __str__(self): return ( f"{', '.join([str(c) for c in self._constraints[:-1]])} or" f" {self._constraints[-1]}" ) class _MissingValues(_Constraint): """Helper constraint for the `missing_values` parameters. Convenience for [ Integral, Interval(Real, None, None, closed="both"), str, None, _NanConstraint(), _PandasNAConstraint(), ] """ def __init__(self): super().__init__() self._constraints = [ _InstancesOf(Integral), # we use an interval of Real to ignore np.nan that has its own constraint Interval(Real, None, None, closed="both"), _InstancesOf(str), _NoneConstraint(), _NanConstraint(), _PandasNAConstraint(), ] def is_satisfied_by(self, val): return any(c.is_satisfied_by(val) for c in self._constraints) def __str__(self): return ( f"{', '.join([str(c) for c in self._constraints[:-1]])} or" f" {self._constraints[-1]}" ) class HasMethods(_Constraint): """Constraint representing objects that expose specific methods. It is useful for parameters following a protocol and where we don't want to impose an affiliation to a specific module or class. Parameters ---------- methods : str or list of str The method(s) that the object is expected to expose. """ @validate_params({"methods": [str, list]}) def __init__(self, methods): super().__init__() if isinstance(methods, str): methods = [methods] self.methods = methods def is_satisfied_by(self, val): return all(callable(getattr(val, method, None)) for method in self.methods) def __str__(self): if len(self.methods) == 1: methods = f"{self.methods[0]!r}" else: methods = ( f"{', '.join([repr(m) for m in self.methods[:-1]])} and" f" {self.methods[-1]!r}" ) return f"an object implementing {methods}" class _IterablesNotString(_Constraint): """Constraint representing iterables that are not strings.""" def is_satisfied_by(self, val): return isinstance(val, Iterable) and not isinstance(val, str) def __str__(self): return "an iterable" class _CVObjects(_Constraint): """Constraint representing cv objects. Convenient class for [ Interval(Integral, 2, None, closed="left"), HasMethods(["split", "get_n_splits"]), _IterablesNotString(), None, ] """ def __init__(self): super().__init__() self._constraints = [ Interval(Integral, 2, None, closed="left"), HasMethods(["split", "get_n_splits"]), _IterablesNotString(), _NoneConstraint(), ] def is_satisfied_by(self, val): return any(c.is_satisfied_by(val) for c in self._constraints) def __str__(self): return ( f"{', '.join([str(c) for c in self._constraints[:-1]])} or" f" {self._constraints[-1]}" ) class Hidden: """Class encapsulating a constraint not meant to be exposed to the user. Parameters ---------- constraint : str or _Constraint instance The constraint to be used internally. """ def __init__(self, constraint): self.constraint = constraint def generate_invalid_param_val(constraint, constraints=None): """Return a value that does not satisfy the constraint. Raises a NotImplementedError if there exists no invalid value for this constraint. This is only useful for testing purpose. Parameters ---------- constraint : _Constraint instance The constraint to generate a value for. constraints : list of _Constraint instances or None, default=None The list of all constraints for this parameter. If None, the list only containing `constraint` is used. Returns ------- val : object A value that does not satisfy the constraint. """ if isinstance(constraint, StrOptions): return f"not {' or '.join(constraint.options)}" if isinstance(constraint, _MissingValues): return np.array([1, 2, 3]) if isinstance(constraint, _VerboseHelper): return -1 if isinstance(constraint, HasMethods): return type("HasNotMethods", (), {})() if isinstance(constraint, _IterablesNotString): return "a string" if isinstance(constraint, _CVObjects): return "not a cv object" if not isinstance(constraint, Interval): raise NotImplementedError # constraint is an interval constraints = [constraint] if constraints is None else constraints return _generate_invalid_param_val_interval(constraint, constraints) def _generate_invalid_param_val_interval(interval, constraints): """Return a value that does not satisfy an interval constraint. Generating an invalid value for an integer interval depends on the other constraints since an int is a real, meaning that it can be valid for a real interval. Assumes that there can be at most 2 interval constraints: one integer interval and/or one real interval. This is only useful for testing purpose. Parameters ---------- interval : Interval instance The interval to generate a value for. constraints : list of _Constraint instances The list of all constraints for this parameter. Returns ------- val : object A value that does not satisfy the interval constraint. """ if interval.type is Real: # generate a non-integer value such that it can't be valid even if there's also # an integer interval constraint. if interval.left is None and interval.right is None: if interval.closed in ("left", "neither"): return np.inf elif interval.closed in ("right", "neither"): return -np.inf else: raise NotImplementedError if interval.left is not None: return np.floor(interval.left) - 0.5 else: # right is not None return np.ceil(interval.right) + 0.5 else: # interval.type is Integral if interval.left is None and interval.right is None: raise NotImplementedError # We need to check if there's also a real interval constraint to generate a # value that is not valid for any of the 2 interval constraints. real_intervals = [ i for i in constraints if isinstance(i, Interval) and i.type is Real ] real_interval = real_intervals[0] if real_intervals else None if real_interval is None: # Only the integer interval constraint -> easy if interval.left is not None: return interval.left - 1 else: # interval.right is not None return interval.right + 1 # There's also a real interval constraint. Try to find a value left to both or # right to both or in between them. # redefine left and right bounds to be smallest and largest valid integers in # both intervals. int_left = interval.left if int_left is not None and interval.closed in ("right", "neither"): int_left = int_left + 1 int_right = interval.right if int_right is not None and interval.closed in ("left", "neither"): int_right = int_right - 1 real_left = real_interval.left if real_interval.left is not None: real_left = int(np.ceil(real_interval.left)) if real_interval.closed in ("right", "neither"): real_left = real_left + 1 real_right = real_interval.right if real_interval.right is not None: real_right = int(np.floor(real_interval.right)) if real_interval.closed in ("left", "neither"): real_right = real_right - 1 if int_left is not None and real_left is not None: # there exists an int left to both intervals return min(int_left, real_left) - 1 if int_right is not None and real_right is not None: # there exists an int right to both intervals return max(int_right, real_right) + 1 if int_left is not None: if real_right is not None and int_left - real_right >= 2: # there exists an int between the 2 intervals return int_left - 1 else: raise NotImplementedError else: # int_right is not None if real_left is not None and real_left - int_right >= 2: # there exists an int between the 2 intervals return int_right + 1 else: raise NotImplementedError def generate_valid_param(constraint): """Return a value that does satisfy a constraint. This is only useful for testing purpose. Parameters ---------- constraint : Constraint instance The constraint to generate a value for. Returns ------- val : object A value that does satisfy the constraint. """ if isinstance(constraint, _ArrayLikes): return np.array([1, 2, 3]) if isinstance(constraint, _SparseMatrices): return csr_matrix([[0, 1], [1, 0]]) if isinstance(constraint, _RandomStates): return np.random.RandomState(42) if isinstance(constraint, _Callables): return lambda x: x if isinstance(constraint, _NoneConstraint): return None if isinstance(constraint, _InstancesOf): return constraint.type() if isinstance(constraint, _Booleans): return True if isinstance(constraint, _VerboseHelper): return 1 if isinstance(constraint, _MissingValues): return np.nan if isinstance(constraint, HasMethods): return type( "ValidHasMethods", (), {m: lambda self: None for m in constraint.methods} )() if isinstance(constraint, _IterablesNotString): return [1, 2, 3] if isinstance(constraint, _CVObjects): return 5 if isinstance(constraint, Options): # includes StrOptions for option in constraint.options: return option if isinstance(constraint, Interval): interval = constraint if interval.left is None and interval.right is None: return 0 elif interval.left is None: return interval.right - 1 elif interval.right is None: return interval.left + 1 else: if interval.type is Real: return (interval.left + interval.right) / 2 else: return interval.left + 1 raise ValueError(f"Unknown constraint type: {constraint}")