"""Define Parameter which holds the value."""
# Copyright (c) 2024 zfit
from __future__ import annotations
import abc
import collections
import copy
import functools
import typing
import warnings
import weakref
from collections.abc import Callable, Iterable
from contextlib import suppress
from inspect import signature
from typing import Literal, Mapping, Optional, Union
from weakref import WeakSet
import dill
import numpy as np
import pydantic
import tensorflow as tf
import tensorflow_probability as tfp
# TF backwards compatibility
from ordered_set import OrderedSet
from pydantic import Field, validator
from tensorflow.python.ops import tensor_getitem_override
from tensorflow.python.ops.resource_variable_ops import (
ResourceVariable as TFVariable,
)
from tensorflow.python.ops.resource_variable_ops import (
VariableSpec,
)
from tensorflow.python.ops.variables import Variable
from tensorflow.python.types.core import Tensor as TensorType
import zfit.z.numpy as znp
from .. import z
from ..core.baseobject import BaseNumeric, extract_filter_params
from ..minimizers.interface import ZfitResult
from ..serialization.paramrepr import make_param_constructor
from ..serialization.serializer import BaseRepr, Serializer
from ..settings import run, ztypes
from ..util import ztyping
from ..util.checks import NotSpecified
from ..util.container import convert_to_container
from ..util.deprecation import deprecated, deprecated_args
from ..util.exception import (
BreakingAPIChangeError,
FunctionNotImplemented,
IllegalInGraphModeError,
LogicalUndefinedOperationError,
ParameterNotIndependentError,
)
from ..util.temporary import TemporarilySet
from . import interfaces as zinterfaces
from .dependents import _extract_dependencies
from .interfaces import ZfitIndependentParameter, ZfitModel, ZfitParameter
from .serialmixin import SerializableMixin
# todo add type hints in this module for api
class MetaBaseParameter(type(tf.Variable), type(zinterfaces.ZfitParameter)): # resolve metaclasses
pass
def register_tensor_conversion(
convertable, name=None, overload_operators=True, priority=10
): # higher than any tf conversion
def _dense_var_to_tensor(var, dtype=None, name=None, as_ref=False):
return var._dense_var_to_tensor(dtype=dtype, name=name, as_ref=as_ref)
tf.register_tensor_conversion_function(convertable, _dense_var_to_tensor, priority=priority)
if name:
pass
# _pywrap_utils.RegisterType(name, convertable)
if overload_operators:
convertable._OverloadAllOperators()
class OverloadableMixin(ZfitParameter):
# Conversion to tensor.
@staticmethod
def _TensorConversionFunction(v, dtype=None, name=None, as_ref=False): # pylint: disable=invalid-name
"""Utility function for converting a Variable to a Tensor."""
_ = name
if dtype and not dtype.is_compatible_with(v.dtype):
msg = (
f"Incompatible type conversion requested to type '{dtype.name}' for variable "
f"of type '{v.dtype.name}'"
)
raise ValueError(msg)
if as_ref:
return v._ref() # pylint: disable=protected-access
else:
return v.value()
def _dense_var_to_tensor(self, dtype=None, name=None, as_ref=False):
del name
if dtype and not dtype.is_compatible_with(self.dtype):
msg = (
f"Incompatible type conversion requested to type '{dtype.name}' for variable "
f"of type '{self.dtype.name}'"
)
raise ValueError(msg)
if as_ref:
if hasattr(self, "_ref"):
return self._ref()
else:
msg = "Why is this needed?"
raise RuntimeError(msg)
else:
return self.value()
def _AsTensor(self):
return self.value()
@classmethod
def _OverloadAllOperators(cls): # pylint: disable=invalid-name
"""Register overloads for all operators."""
for operator in tf.Tensor.OVERLOADABLE_OPERATORS:
cls._OverloadOperator(operator)
# For slicing, bind getitem differently than a tensor (use SliceHelperVar
# instead)
# pylint: disable=protected-access
cls.__getitem__ = tensor_getitem_override._slice_helper_var
@classmethod
def _OverloadOperator(cls, operator): # pylint: disable=invalid-name
"""Defer an operator overload to ``ops.Tensor``.
We pull the operator out of ops.Tensor dynamically to avoid ordering issues.
Args:
operator: string. The operator name.
"""
# We can't use the overload mechanism on __eq__ & __ne__ since __eq__ is
# called when adding a variable to sets. As a result we call a.value() which
# causes infinite recursion when operating within a GradientTape
# TODO(gjn): Consider removing this
if operator in ("__eq__", "__ne__"):
return
tensor_oper = getattr(tf.Tensor, operator)
def _run_op(a, *args, **kwargs):
# pylint: disable=protected-access
return tensor_oper(a.value(), *args, **kwargs)
functools.update_wrapper(_run_op, tensor_oper)
setattr(cls, operator, _run_op)
register_tensor_conversion(OverloadableMixin, overload_operators=True)
class WrappedVariable(metaclass=MetaBaseParameter):
def __init__(self, initial_value, *args, **kwargs):
super().__init__(*args, **kwargs)
self.variable = tf.Variable(
initial_value=initial_value,
name=self.name,
dtype=self.dtype,
)
@property
@abc.abstractmethod
def name(self):
raise NotImplementedError
@property
def constraint(self):
return self.variable.constraint
@property
def dtype(self):
return self.variable.dtype
def value(self):
return self.variable.value()
def read_value(self):
return self.variable.read_value()
@property
def shape(self):
return self.variable.shape
def numpy(self):
return self.variable.numpy()
def assign(self, value, use_locking=False, name=None, read_value=True):
return self.variable.assign(value=value, use_locking=use_locking, name=name, read_value=read_value)
def _dense_var_to_tensor(self, dtype=None, name=None, as_ref=False):
del name
if dtype is not None and dtype != self.dtype:
return NotImplemented
if as_ref:
return self.variable.read_value().op.inputs[0]
else:
return self.variable.value()
def _AsTensor(self):
return self.variable.value()
@staticmethod
def _OverloadAllOperators(): # pylint: disable=invalid-name
"""Register overloads for all operators."""
for operator in tf.Tensor.OVERLOADABLE_OPERATORS:
WrappedVariable._OverloadOperator(operator)
# For slicing, bind getitem differently than a tensor (use SliceHelperVar
# instead)
# pylint: disable=protected-access
WrappedVariable.__getitem__ = tensor_getitem_override._slice_helper_var
@staticmethod
def _OverloadOperator(operator): # pylint: disable=invalid-name
"""Defer an operator overload to ``ops.Tensor``.
We pull the operator out of ops.Tensor dynamically to avoid ordering issues.
Args:
operator: string. The operator name.
"""
tensor_oper = getattr(tf.Tensor, operator)
def _run_op(a, *args):
# pylint: disable=protected-access
value = a._AsTensor()
return tensor_oper(value, *args)
# Propagate __doc__ to wrapper
with suppress(AttributeError):
_run_op.__doc__ = tensor_oper.__doc__
setattr(WrappedVariable, operator, _run_op)
register_tensor_conversion(WrappedVariable, "WrappedVariable", overload_operators=True)
class BaseParameter(Variable, ZfitParameter, TensorType, metaclass=MetaBaseParameter):
def __init__(self, *args, **kwargs):
try:
super().__init__(*args, **kwargs)
except NotImplementedError as error:
tmp_val = kwargs.pop("name", None) # remove if name is in there, needs to be passed through
if args or kwargs:
kwargs["name"] = tmp_val
msg = (
f"The following arguments reached the top of the inheritance tree, the super "
f"init is not implemented (most probably abstract tf.Variable): {args, kwargs}. "
f"If you see this error, please post it as an bug at: "
f"https://github.com/zfit/zfit/issues/new/choose"
)
raise RuntimeError(msg) from error
def __len__(self):
return 1
class ZfitParameterMixin(BaseNumeric):
_existing_params: typing.ClassVar = {}
def __init__(self, name, label=None, **kwargs):
if name not in self._existing_params:
self._existing_params[name] = WeakSet()
# Is an alternative arg for pop needed in case it fails? Why would it fail?
weakref.finalize(
self,
lambda name: name in self._existing_params and self._existing_params.pop,
name,
)
self._existing_params[name].add(self)
self._name = name
self._label = label
super().__init__(name=name, **kwargs)
self._assert_params_unique()
# property needed here to overwrite the name of tf.Variable
@property
def name(self) -> str:
return self._name
@property
def label(self) -> str:
if (label := self._label) is None:
label = self.name
return label
def __del__(self):
with suppress(AttributeError, NotImplementedError): # if super does not have a __del__
super().__del__(self)
def __add__(self, other):
if isinstance(other, (ZfitModel, ZfitParameter)):
from . import operations
with suppress(FunctionNotImplemented):
return operations.add(self, other)
return super().__add__(other)
def __radd__(self, other):
if isinstance(other, (ZfitModel, ZfitParameter)):
from . import operations
with suppress(FunctionNotImplemented):
return operations.add(other, self)
return super().__radd__(other)
def __mul__(self, other):
if isinstance(other, (ZfitModel, ZfitParameter)):
from . import operations
with suppress(FunctionNotImplemented):
return operations.multiply(self, other)
return super().__mul__(other)
def __rmul__(self, other):
if isinstance(other, (ZfitModel, ZfitParameter)):
from . import operations
with suppress(FunctionNotImplemented):
return operations.multiply(other, self)
return super().__rmul__(other)
def __eq__(self, other):
return self is other
def __hash__(self):
if not hasattr(self, "_cached_hash"):
self._cached_hash = hash(id(self))
return self._cached_hash
class TFBaseVariable(TFVariable, metaclass=MetaBaseParameter):
# class TFBaseVariable(WrappedVariable, metaclass=MetaBaseParameter):
# Needed, otherwise tf variable complains about the name not having a ':' in there
@property
def _shared_name(self):
return self.name
[docs]
class Parameter(
ZfitParameterMixin,
TFBaseVariable,
BaseParameter,
SerializableMixin,
ZfitIndependentParameter,
):
"""Class for fit parameters that has a default state."""
_independent = True
_independent_params = WeakSet()
DEFAULT_STEP_SIZE = 0.01
def __init__(
self,
name: str,
value: ztyping.NumericalScalarType,
lower: ztyping.NumericalScalarType | None = None,
upper: ztyping.NumericalScalarType | None = None,
step_size: ztyping.NumericalScalarType | None = None,
floating: bool = True,
*,
dtype: tf.DType = None,
label: str | None = None,
# legacy
lower_limit: ztyping.NumericalScalarType | None = None,
upper_limit: ztyping.NumericalScalarType | None = None,
):
"""Fit Parameter that has a default state (value) and limits (lower, upper).
The name identifies the parameter.
Multiple parameters with the same name can exist, however,
they cannot be in the same PDF/func/loss as the value would not be uniquely defined.
Args:
name : Name of the parameter. Should be unique within a model/likelihood.
value : Default value of the parameter. Also used as the starting value in minimization.
lower : lower limit of the parameter. If the parameter is set to a value below the lower limit, it will raise an error.
upper : upper limit of the parameter. If the parameter is set to a value above the upper limit, it will raise an error.
floating : If the parameter is floating (can change value) or fixed (constant) in the minimization.
label: |@doc:param.init.label||@docend:param.init.label|
step_size : Initial step size for minimization. If not set, a default value is used.
"""
self._independent_params.add(self)
# legacy start
if lower_limit is not None:
msg = "The argument `lower_limit` has been renamed to `lower`."
raise BreakingAPIChangeError(msg)
if upper_limit is not None:
msg = "The argument `upper_limit` has been renamed to `upper`."
raise BreakingAPIChangeError(msg)
if dtype is None:
dtype = ztypes.float
else:
warnings.warn(
"The argument `dtype` is deprecated and will be removed in the future.",
DeprecationWarning,
stacklevel=2,
)
# legacy end
# TODO: sanitize input for TF2
self._lower_limit_neg_inf = None
self._upper_limit_neg_inf = None
if lower is None:
self._lower_limit_neg_inf = znp.asarray(-np.inf, dtype)
if upper is None:
self._upper_limit_neg_inf = znp.asarray(np.inf, dtype)
value = znp.asarray(value, dtype=ztypes.float)
super().__init__(
initial_value=value,
dtype=dtype,
name=name,
params={},
label=label,
)
self.lower = lower
self.upper = upper
self.floating = floating
self.step_size = step_size
self.set_value(value) # to check that it is in the limits
def __init_subclass__(cls, **kwargs):
super().__init_subclass__(**kwargs)
cls._independent = True # overwriting independent only for subclass/instance
@classmethod
def _from_name(cls, name):
for param in cls._independent_params:
if name == param.name:
return param
msg = f"Parameter {name} does not exist, please create it first."
raise ValueError(msg)
@property
def lower(self):
limit = self._lower
if limit is None:
limit = self._lower_limit_neg_inf
return limit
@lower.setter
# @invalidate_graph
def lower(self, value):
if value is None and self._lower_limit_neg_inf is None:
self._lower_limit_neg_inf = znp.asarray(-np.inf, dtype=ztypes.float)
elif value is not None:
value = znp.asarray(value, dtype=ztypes.float)
self._lower = value
@property
def upper(self):
limit = self._upper
if limit is None:
limit = self._upper_limit_neg_inf
return limit
@upper.setter
# @invalidate_graph
def upper(self, value):
if value is None and self._upper_limit_neg_inf is None:
self._upper_limit_neg_inf = znp.asarray(np.inf, dtype=ztypes.float)
elif value is not None:
value = znp.asarray(value, dtype=ztypes.float)
self._upper = value
@property
def has_limits(self) -> bool:
"""If the parameter has limits set or not."""
no_limits = self._lower is None and self._upper is None
return not no_limits
@property
def at_limit(self) -> tf.Tensor:
"""If the value is at the limit (or over it).
Returns:
Boolean ``tf.Tensor`` that tells whether the value is at the limits.
"""
return self._check_at_limit(self.value())
def _check_at_limit(self, value, exact=False):
"""The precision is up to 1e-5 relative or 1e-8 absolute if exact is None.
Args:
value ():
exact ():
Returns:
"""
if not self.has_limits:
return tf.constant(False)
# Adding a slight tolerance to make sure we're not tricked by numerics due to floating point comparison
diff = znp.abs(self.upper - self.lower) # catch if it is minus inf
if not exact:
reltol = 0.005
abstol = 1e-5
else:
reltol = 1e-5
abstol = 1e-7
tol = znp.minimum(diff * reltol, abstol) # if one limit is inf we would get inf
if not exact: # if exact, we wanna allow to set it slightly over the limit.
tol = -tol # If not, we wanna make sure it's inside
at_lower = z.unstable.less_equal(value, self.lower - tol)
at_upper = z.unstable.greater_equal(value, self.upper + tol)
return z.unstable.logical_or(at_lower, at_upper)
def value(self):
value = super().value()
if self.has_limits:
value = tfp.math.clip_by_value_preserve_gradient(
value, clip_value_min=self.lower, clip_value_max=self.upper
)
return value
[docs]
@deprecated(None, "Use `value` instead.")
def read_value(self):
return self.value()
@property
def floating(self):
if self._floating and (hasattr(self, "trainable") and not self.trainable):
msg = "Floating is set to true but tf Variable is not trainable."
raise RuntimeError(msg)
return self._floating
@floating.setter
def floating(self, value):
if not isinstance(value, bool):
msg = "floating has to be a boolean."
raise TypeError(msg)
self._floating = value
def _get_dependencies(self):
return {self}
@property
def independent(self):
return self._independent
@property
def has_step_size(self):
return self._step_size is not None
@property
def step_size(self) -> tf.Tensor: # TODO: improve default step_size?
"""Step size of the parameter, the estimated order of magnitude of the uncertainty.
This can be crucial to tune for the minimization. A too large ``step_size`` can produce NaNs, a too small won't
converge.
If the step size is not set, the ``DEFAULT_STEP_SIZE`` is used.
Returns:
The step size
"""
step_size = self._step_size
if step_size is None:
# # auto-infer from limits
# step_splits = 1e5
# if self.has_limits:
# step_size = (self.upper_limit - self.lower_limit) / step_splits # TODO improve? can be tensor?
# else:
# step_size = self.DEFAULT_STEP_SIZE
# if np.isnan(step_size):
# if self.lower_limit == -np.infty or self.upper_limit == np.infty:
# step_size = self.DEFAULT_STEP_SIZE
# else:
# raise ValueError("Could not set step size. Is NaN.")
# # step_size = z.to_real(step_size)
# self.step_size = step_size
step_size = self.DEFAULT_STEP_SIZE
# step_size = z.convert_to_tensor(step_size)
return step_size
@step_size.setter
def step_size(self, value):
if value is not None:
value = float(value)
# value = z.convert_to_tensor(value, preferred_dtype=ztypes.float)
# value = znp.asarray(value, dtype=ztypes.float)
self._step_size = value
[docs]
def set_value(self, value: ztyping.NumericalScalarType):
"""Set the :py:class:`~zfit.Parameter` to `value` (temporarily if used in a context manager).
This operation won't, compared to the assign, return the read value but an object that *can* act as a context
manager.
Args:
value: The value the parameter will take on.
Raises:
ValueError: If the value is not inside the limits (in normal Python/eager mode)
InvalidArgumentError: If the value is not inside the limits (in JIT/traced/graph mode)
"""
def getter():
return self.value()
def setter(value):
if self.has_limits:
message = (
f"Setting value {value} invalid for parameter {self.name} with limits "
f"{self.lower} - {self.upper}. This is changed."
f" In order to silence this and clip the value, you can use (with caution,"
f" advanced) `Parameter.assign`"
)
if run.executing_eagerly():
if self._check_at_limit(value, exact=True):
raise ValueError(message)
else:
tf.debugging.assert_greater(
znp.asarray(value, tf.float64),
znp.asarray(self.lower, tf.float64),
message=message,
)
tf.debugging.assert_less(
znp.asarray(value, tf.float64),
znp.asarray(self.upper, tf.float64),
message=message,
)
# tf.debugging.Assert(self._check_at_limit(value), [value])
self.assign(value=value, read_value=False)
return TemporarilySet(value=value, setter=setter, getter=getter)
[docs]
def assign(self, value, use_locking: bool | None = None, read_value=False):
"""Set the :py:class:`~zfit.Parameter` to `value` without any checks.
Compared to ``set_value``, this method cannot be used with a context manager and won't raise an
error
Args:
value: The value the parameter will take on.
"""
return super().assign(value=value, use_locking=use_locking, read_value=read_value)
[docs]
def randomize(
self,
minval: ztyping.NumericalScalarType | None = None,
maxval: ztyping.NumericalScalarType | None = None,
sampler: Callable = np.random.uniform,
) -> tf.Tensor:
"""Update the parameter with a randomised value between minval and maxval and return it.
Args:
minval: The lower bound of the sampler. If not given, ``lower_limit`` is used.
maxval: The upper bound of the sampler. If not given, ``upper_limit`` is used.
sampler: A sampler with the same interface as ``np.random.uniform``
Returns:
The sampled value
"""
if not tf.executing_eagerly():
msg = "Randomizing values in a parameter within Graph mode is most probably not" " what is "
raise IllegalInGraphModeError(msg)
minval = self.lower if minval is None else znp.asarray(minval, dtype=self.dtype)
maxval = self.upper if maxval is None else znp.asarray(maxval, dtype=self.dtype)
if maxval is None or minval is None:
msg = "Cannot randomize a parameter without limits or limits given."
raise RuntimeError(msg)
value = sampler(size=self.shape, low=minval, high=maxval)
self.set_value(value=value)
return value
def get_params(
self,
floating: bool | None = True,
is_yield: bool | None = None,
extract_independent: bool | None = True,
only_floating=NotSpecified,
) -> set[ZfitParameter]:
del is_yield, only_floating, extract_independent # doesnot make sense for a single parameter
return extract_filter_params(self, floating=floating, extract_independent=False)
def __repr__(self): # many try and except in case it's not fully initialized yet
if tf.executing_eagerly(): # more explicit: we check for exactly this attribute, nothing inside numpy
try:
value = f"{self.numpy():.4g}"
except Exception as err:
value = f"errored {err}"
else:
value = "graph-node"
try:
floating = self.floating
except Exception as err:
floating = f"errored {err}"
try:
name = self.name
except Exception as err:
name = f"errored {err}"
return f"<zfit.{self.__class__.__name__} '{name}' floating={floating} value={value}>"
# LEGACY, deprecate?
@property
@deprecated(None, "Use `lower` instead.")
def lower_limit(self):
return self.lower
@lower_limit.setter
@deprecated(None, "Use `lower` instead.")
def lower_limit(self, value):
self.lower = value
@property
@deprecated(None, "Use `upper` instead.")
def upper_limit(self):
return self.upper
@upper_limit.setter
@deprecated(None, "Use `upper` instead.")
def upper_limit(self, value):
self.upper = value
def __tf_tracing_type__(self, signature_context):
return ParameterSpec(parameter=self)
# below needed?
# CompositeTensor method
# @property
# def _type_spec(self):
# return ParameterSpec(parameter=self)
#
# # CompositeTensor method
# def _shape_invariant_to_type_spec(self, shape):
# return ParameterSpec(shape, self.dtype, self.trainable)
def __reduce__(self): # for pickling
return functools.partial(
Parameter,
name=self.name,
value=self.value(),
lower=self.lower,
upper=self.upper,
floating=self.floating,
label=self.label,
step_size=self.step_size,
), ()
# delattr(Parameter, "__tf_tracing_type__")
class ParameterSpec(VariableSpec):
value_type = property(lambda _: Parameter)
def __init__(self, shape=None, dtype=None, trainable=True, alias_id=None, *, parameter=None):
if parameter is None:
msg = "Unknown error, please report, parameter should not be none"
raise RuntimeError(msg)
if parameter is not None: # initialize from parameter
shape = parameter.shape
dtype = parameter.dtype
trainable = True
alias_id = None
self.parameter_value = parameter
self._name = parameter.name
self.parameter_type = type(parameter)
if dtype is None:
dtype = tf.float64
super().__init__(shape=shape, dtype=dtype, trainable=trainable, alias_id=alias_id)
self.hash = hash(id(parameter))
@classmethod
def from_value(cls, value):
return cls(parameter=value)
def _to_components(self, value):
return super()._to_components(value)
def _from_components(self, components):
_ = super()._from_components(components) # checking that there is no error
return Parameter._from_name(self.name)
def _to_tensors(self, value):
return [value]
def is_subtype_of(self, other):
return (
type(other) is ParameterSpec
and self.parameter_type is other.parameter_type
and self.parameter_value is other.parameter_value
)
def most_specific_common_supertype(self, others):
return self if all(self == other for other in others) else None
def placeholder_value(self, placeholder_context=None):
del placeholder_context # unused
return self.parameter_value
def is_compatible_with(self, spec_or_value):
return isinstance(spec_or_value, ParameterSpec) and self.parameter_value is spec_or_value.parameter_value
def __eq__(self, other) -> bool:
if not isinstance(other, ParameterSpec):
return False
return self.parameter_value is other.parameter_value
def __hash__(self):
return self.hash
class ParameterRepr(BaseRepr): # add label?
_implementation = Parameter
_constructor = pydantic.PrivateAttr(make_param_constructor(Parameter))
hs3_type: Literal["Parameter"] = Field("Parameter", alias="type")
name: str
value: float
lower: Optional[float] = Field(None, alias="min")
upper: typing.Optional[float] = Field(None, alias="max")
step_size: Optional[float] = None
floating: Optional[bool] = None
@validator("value", pre=True)
def _validate_value(cls, v):
if cls.orm_mode(v):
v = v()
return v
def __hash__(self):
return hash(self.name) + hash(type(self))
class BaseComposedParameter(ZfitParameterMixin, OverloadableMixin, BaseParameter):
def __init__(self, params, func, dtype=None, name="BaseComposedParameter", **kwargs):
# 0.4 breaking
if "value" in kwargs:
msg = "'value' cannot be provided any longer, `func` is needed."
raise BreakingAPIChangeError(msg)
super().__init__(name=name, params=params, **kwargs)
if not callable(func):
msg = "`func` is not callable."
raise TypeError(msg)
self._func = func
self._dtype = dtype if dtype is not None else ztypes.float
def _get_dependencies(self):
return _extract_dependencies(list(self.params.values()))
@property
def floating(self):
msg = "Cannot be floating or not. Look at the dependencies."
raise LogicalUndefinedOperationError(msg)
@floating.setter
def floating(self, value): # noqa: ARG002
msg = "Cannot set floating or not. Set in the dependencies (`get_params`)."
raise LogicalUndefinedOperationError(msg)
@property
def params(self):
return self._params
def value(self):
params = self.params
return znp.asarray(self._func(params), dtype=self.dtype)
# return tf.convert_to_tensor(value, dtype=self.dtype)
def read_value(self):
return tf.identity(self.value())
@property
def shape(self):
return self.value().shape
def numpy(self):
return self.value().numpy()
@property
def independent(self):
return False
def set_value(self, value): # noqa: ARG002
"""Set the value of the parameter. Cannot be used for composed parameters!
Args:
value:
"""
msg = "Cannot set value of a composed parameter." " Set the value on its components."
raise LogicalUndefinedOperationError(msg)
def randomize(self, minval=None, maxval=None, sampler=np.random.uniform): # noqa: ARG002
"""Randomize the value of the parameter.
Cannot be used for composed parameters!
"""
msg = "Cannot randomize a composed parameter."
raise LogicalUndefinedOperationError(msg)
def assign(self, value, use_locking=False, name=None, read_value=True): # noqa: ARG002
"""Assign the value of the parameter.
Cannot be used for composed parameters!
"""
msg = "Cannot assign a composed parameter."
raise LogicalUndefinedOperationError(msg)
[docs]
class ConstantParameter(OverloadableMixin, ZfitParameterMixin, BaseParameter, SerializableMixin):
"""Constant parameter.
Value cannot change.
"""
def __init__(
self,
name,
value,
*,
label: str | None = None,
):
"""Constant parameter that cannot change its value.
Args:
name: Unique identifier of the parameter.
value: Constant value.
label: |@doc:param.init.label||@docend:param.init.label|
"""
super().__init__(name=name, params={}, dtype=ztypes.float, label=label)
self._value_np = tf.get_static_value(value, partial=True)
self._value = tf.guarantee_const(tf.convert_to_tensor(value, dtype=self.dtype))
@property
def shape(self):
return self.value().shape
def value(self) -> tf.Tensor:
return self._value
def read_value(self) -> tf.Tensor:
return self.value()
@property
def floating(self):
return False
@floating.setter
def floating(self, value): # noqa: ARG002
msg = "Cannot set a ConstantParameter to floating. Use a `Parameter` instead."
raise LogicalUndefinedOperationError(msg)
@property
def independent(self) -> bool:
return False
def _get_dependencies(self) -> ztyping.DependentsType:
return OrderedSet()
@property
def static_value(self):
return self._value_np
def numpy(self):
return self._value_np
def __repr__(self):
value_str = f"{value: .4g}" if (value := self._value_np) is not None else "symbolic"
return f"<zfit.param.{self.__class__.__name__} '{self.name}' dtype={self.dtype.name} value={value_str}>"
register_tensor_conversion(ConstantParameter, "ConstantParameter", overload_operators=True)
register_tensor_conversion(BaseComposedParameter, "BaseComposedParameter", overload_operators=True)
class ConstantParamRepr(BaseRepr):
_implementation = ConstantParameter
_constructor = pydantic.PrivateAttr(make_param_constructor(ConstantParameter))
hs3_type: Literal["ConstantParameter"] = pydantic.Field("ConstantParameter", alias="type")
name: str
value: float
floating: bool = False
# lower: Optional[float] = Field(None, alias="min")
# upper: Optional[float] = Field(None, alias="max")
@validator("value", pre=True)
def _validate_value(cls, value):
if cls.orm_mode(value):
value = value()
return value
def _to_orm(self, init):
init = copy.copy(init)
init.pop("floating")
return super()._to_orm(init)
[docs]
class ComposedParameter(SerializableMixin, BaseComposedParameter):
@deprecated_args(None, "Use `params` instead.", "dependents")
@deprecated_args(None, "Use `func` instead.", "value_fn")
def __init__(
self,
name: str,
value_fn: Optional[Callable] = None,
*,
func: Optional[Callable] = None,
params: (dict[str, ZfitParameter] | Iterable[ZfitParameter] | ZfitParameter) = NotSpecified,
dtype: tf.dtypes.DType = ztypes.float,
label: str | None = None,
unpack_params: bool | None = None,
dependents: (dict[str, ZfitParameter] | Iterable[ZfitParameter] | ZfitParameter) = NotSpecified,
):
"""Arbitrary composition of parameters.
A `ComposedParameter` allows for arbitrary combinations of parameters and correlations using an arbitrary
function.
Examples:
.. jupyter-execute::
import zfit
param1 = zfit.Parameter('param1', 1.0, 0.1, 1.8)
param2 = zfit.Parameter('param2', 42.0, 0, 100)
# using a dict for the params
def mult_dict(params):
return params["a"] * params["b"]
mult_param_dict = zfit.ComposedParameter('mult_dict', mult_dict, params={"a": param1, "b": param2})
# using a list for the params
def mult_list(params):
return params[0] * params[1]
mult_param_list = zfit.ComposedParameter('mult_list', mult_list, params=[param1, param2])
Args:
name: Unique name of the Parameter.
func: Function that returns the value of the composed parameter and takes as arguments `params` as
arguments. The function must be able to be called with the same arguments as `params`.
params: If it is a `dict`, this will directly be used as the `params` attribute, otherwise the
parameters will be automatically named with f"param_{i}". The values act as arguments to `func`.
dtype: Output of `func` dtype
label: |@doc:param.init.label||@docend:param.init.label|
unpack_params: If True, the parameters will be unpacked and passed as arguments to `func`. If False, the
parameters will be passed as a dict/tuple. If None, it will be automatically determined and raise an error
if it cannot be determined.
dependents:
.. deprecated:: unknown
use `params` instead.
"""
# legacy
if value_fn is not None:
if func is not None:
msg = "Cannot specify both `value_fn` and `func`."
raise ValueError(msg)
func = value_fn
del value_fn
# end legacy
if not isinstance(params, Mapping):
params = convert_to_container(params)
original_init = {"name": name, "internal_params": params, "func": func, "unpack_params": unpack_params}
# legacy
if dependents is not NotSpecified:
params = dependents
elif params is NotSpecified:
raise ValueError
# end legacy
if params is None:
msg = "Params needs to be specified."
raise BreakingAPIChangeError(msg)
takes_no_params = False
if unpack_params is None:
parameters = signature(func).parameters
if isinstance(params, ZfitParameter):
params = [params]
unpack_params = True
elif len(parameters) == 1 and (len(params) > 1) or "params" in parameters:
unpack_params = False
elif len(parameters) - len(params) >= 0:
unpack_params = True
elif len(parameters) == 0:
takes_no_params = True
unpack_params = False
else:
msg = (
"Cannot determine if parameter should be unpacked or not. Please specify explicitly `unpack_params`"
)
raise ValueError(msg)
dictlike = isinstance(params, collections.abc.Mapping)
if dictlike:
params_dict = params
if takes_no_params:
def stratified_fn(params):
del params
return func()
elif unpack_params:
def stratified_fn(params):
return func(**params)
else:
stratified_fn = func
else:
params = convert_to_container(params)
if params is None or takes_no_params:
params_dict = {}
def stratified_fn(params):
del params
return func()
else:
params_dict = {f"param_{i}": p for i, p in enumerate(params)}
if unpack_params:
def stratified_fn(params):
return func(*tuple(params.values()))
else:
def stratified_fn(params):
return func(tuple(params.values()))
original_init["params"] = params_dict # needs to be here, we need the params to be a dict for the serialization
super().__init__(params=params_dict, func=stratified_fn, name=name, dtype=dtype, label=label)
self.hs3.original_init.update(original_init)
def __repr__(self):
try:
value = f"{self.numpy():.4g}" if tf.executing_eagerly() else "graph-node"
except Exception:
value = "ERROR OCCURRED"
return f"<zfit.{self.__class__.__name__} '{self.name}' params={[(k, p.name) for k, p in self.params.items()]} value={value}>"
class ComposedParameterRepr(BaseRepr):
_implementation = ComposedParameter
_constructor = pydantic.PrivateAttr(make_param_constructor(ComposedParameter))
hs3_type: Literal["ComposedParameter"] = pydantic.Field("ComposedParameter", alias="type")
name: str
func: str
params: dict[str, Serializer.types.ParamTypeDiscriminated]
unpack_params: Optional[bool]
internal_params: Optional[
Union[
Serializer.types.ParamTypeDiscriminated,
list[Serializer.types.ParamTypeDiscriminated],
dict[str, Serializer.types.ParamTypeDiscriminated],
]
]
@validator("func", pre=True)
def _validate_value_pre(cls, value):
if cls.orm_mode(value):
value = dill.dumps(value).hex()
return value
@pydantic.root_validator(pre=True)
def validate_all_functor(cls, values):
if cls.orm_mode(values):
values = values["hs3"].original_init
return values
def _to_orm(self, init):
init = copy.copy(init)
func = init.pop("func")
params = init.pop("internal_params")
init["params"] = params
init["func"] = dill.loads(bytes.fromhex(func))
return super()._to_orm(init)
[docs]
class ComplexParameter(ComposedParameter): # TODO: change to real, imag as input?
@deprecated_args(None, "Use `func` instead.", "value_fn")
def __init__(
self,
name: str,
value_fn: Callable | None = None,
*,
func: Callable | None = None,
params,
dtype=ztypes.complex,
label: str | None = None,
):
"""Create a complex parameter.
.. note::
Use the constructor class methods instead of the __init__() constructor:
- :py:meth:`ComplexParameter.from_cartesian`
- :py:meth:`ComplexParameter.from_polar`
Args:
name: Name of the parameter.
func: Function that returns the value of the complex parameter and takes as arguments the real and
imaginary part.
params: List of the real and imaginary part of the complex parameter.
dtype: Data type of the complex parameter.
label: |@doc:param.init.label||@docend:param.init.label|
"""
# legacy
if value_fn is not None:
if func is not None:
msg = "Cannot specify both `value_fn` and `func`."
raise ValueError(msg)
func = value_fn
del value_fn
super().__init__(name, func=func, params=params, dtype=dtype, label=label)
self._conj = None
self._mod = None
self._arg = None
self._imag = None
self._real = None
[docs]
@classmethod
def from_cartesian(
cls,
name: str,
real: ztyping.NumericalScalarType,
imag: ztyping.NumericalScalarType,
floating: bool = True,
*,
dtype=ztypes.complex,
label: str | None = None,
) -> ComplexParameter: # TODO: correct dtype handling, also below
"""Create a complex parameter from cartesian coordinates.
Args:
name: Name of the parameter.
real: Real part of the complex number.
imag: Imaginary part of the complex number.
floating: If True, the parameter is floating. If False, the parameter is constant.
dtype: Data type of the complex parameter.
label: |@doc:param.init.label||@docend:param.init.label|
"""
real = convert_to_parameter(real, name=name + "_real", prefer_constant=not floating)
imag = convert_to_parameter(imag, name=name + "_imag", prefer_constant=not floating)
param = cls(
name=name,
func=lambda _real, _imag: znp.asarray(tf.complex(_real, _imag), dtype=dtype),
params=[real, imag],
label=label,
)
param._real = real
param._imag = imag
return param
[docs]
@classmethod
def from_polar(
cls,
name: str,
mod: ztyping.NumericalScalarType,
arg: ztyping.NumericalScalarType,
floating=True,
*,
dtype=ztypes.complex,
label: str | None = None,
**__,
) -> ComplexParameter:
"""Create a complex parameter from polar coordinates.
Args:
name: Name of the parameter.
mod: Modulus (r) the complex number.
arg: Argument (phi) of the complex number.
dtype: Data type of the complex parameter.
floating: If True, the parameter is floating. If False, the parameter is constant.
label: |@doc:param.init.label||@docend:param.init.label|
"""
mod = convert_to_parameter(mod, name=name + "_mod", prefer_constant=not floating)
arg = convert_to_parameter(arg, name=name + "_arg", prefer_constant=not floating)
param = cls(
name=name,
func=lambda _mod, _arg: znp.asarray(tf.complex(_mod * znp.cos(_arg), _mod * znp.sin(_arg)), dtype=dtype),
params=[mod, arg],
label=label,
)
param._mod = mod
param._arg = arg
return param
@property
def conj(self):
"""Returns a complex conjugated copy of the complex parameter."""
if self._conj is None:
name = f"{self.name}_conj"
if (label := self._label) is not None:
label = f"Conjugate of {self.label}"
self._conj = ComplexParameter(
name=name,
func=lambda: znp.conj(self),
params=self.get_cache_deps(),
dtype=self.dtype,
label=label,
)
return self._conj
@property
def real(self) -> tf.Tensor:
"""Real part of the complex parameter."""
return znp.real(self)
@property
def imag(self) -> tf.Tensor:
"""Imaginary part of the complex parameter."""
return znp.imag(self)
@property
def mod(self) -> tf.Tensor:
"""Modulus (r) of the complex parameter."""
return znp.abs(self)
@property
def arg(self) -> tf.Tensor:
"""Argument (phi) of the complex parameter."""
return znp.angle(self)
# register_tensor_conversion(ConstantParameter, "ConstantParameter", True)
register_tensor_conversion(ComposedParameter, "ComposedParameter", True)
_auto_number = 0
def get_auto_number():
global _auto_number
auto_number = _auto_number
_auto_number += 1
return auto_number
def _reset_auto_number():
global _auto_number
_auto_number = 0
def convert_to_parameters(
value,
name: str | list[str] | None = None,
prefer_constant: bool | None = None,
lower=None,
upper=None,
step_size=None,
):
if prefer_constant is None:
prefer_constant = True
if isinstance(value, collections.abc.Mapping):
return convert_to_parameters(**value, prefer_constant=False)
value = convert_to_container(value)
is_param_already = [isinstance(val, ZfitIndependentParameter) for val in value]
if all(is_param_already):
return value
elif any(is_param_already):
msg = f"value has to be either ZfitParameters or values, not mixed (currently)." f" Is {value}."
raise ValueError(msg)
params_dict = {
"value": value,
"name": name,
"lower": lower,
"upper": upper,
"step_size": step_size,
}
params_dict = {
key: convert_to_container(val, ignore=np.ndarray) for key, val in params_dict.items() if val is not None
}
lengths = {len(v) for v in params_dict.values()}
if len(lengths) != 1:
msg = f"Inconsistent length in values when converting the parameters: {params_dict}"
raise ValueError(msg)
params = []
for i in range(len(params_dict["value"])):
pdict = {k: params_dict[k][i] for k in params_dict}
params.append(convert_to_parameter(**pdict, prefer_constant=prefer_constant))
return params
[docs]
@deprecated_args(None, "Use `params` instead.", "dependents")
def convert_to_parameter(
value: ztyping.NumericalScalarType | ZfitParameter | Callable,
name: str | None = None,
prefer_constant: bool = True,
params: ZfitParameter | Iterable[ZfitParameter] | None = None,
lower: ztyping.NumericalScalarType | None = None,
upper: ztyping.NumericalScalarType | None = None,
step_size: ztyping.NumericalScalarType | None = None,
*,
label: str | None = None,
# legacy
dependents=None,
) -> ZfitParameter:
"""Convert a *numerical* to a constant/floating parameter or return if already a parameter.
Args:
value: Value of the parameter. If a `ZfitParameter` is passed, it will be returned as is.
name: Name of the parameter. If None, a unique name will be created.
prefer_constant: If True, create a ConstantParameter instead of a Parameter, if possible.
params: If the value is a callable, the parameters that are passed to the callable.
lower: Lower limit of the parameter.
upper: Upper limit of the parameter.
step_size: Step size of the parameter.
label: |@doc:param.init.label||@docend:param.init.label|
"""
# legacy start
if dependents is not None:
params = dependents
# legacy end
if name is not None:
name = str(name)
if callable(value):
if params is None:
msg = "If the value is a callable, the params have to be specified as an empty list/tuple"
raise ValueError(msg)
return ComposedParameter(f"Composed_autoparam_{get_auto_number()}", func=value, params=params, label=label)
if isinstance(value, ZfitParameter): # TODO(Mayou36): autoconvert variable. TF 2.0?
return value
elif isinstance(value, tf.Variable):
msg = f"Currently, cannot autoconvert tf.Variable ({value}) to zfit.Parameter."
raise TypeError(msg)
# convert to Tensor
if not isinstance(value, tf.Tensor):
value = z.to_complex(value) if isinstance(value, complex) else z.to_real(value)
if not run._enable_parameter_autoconversion:
return value
if value.dtype.is_complex:
if name is None:
name = "FIXED_complex_autoparam_" + str(get_auto_number())
if prefer_constant:
complex_params = (
ConstantParameter(name + "_REALPART", value=znp.real(value)),
ConstantParameter(name + "_IMAGPART", value=znp.imag(value)),
)
else:
complex_params = (
Parameter(name + "_REALPART", value=znp.real(value)),
Parameter(name + "_IMAGPART", value=znp.imag(value)),
)
value = ComplexParameter.from_cartesian(name, real=complex_params[0], imag=complex_params[1], label=label)
elif prefer_constant:
if name is None:
name = "FIXED_autoparam_" + str(get_auto_number()) if name is None else name
value = ConstantParameter(name, value=value, label=label)
else:
name = "autoparam_" + str(get_auto_number()) if name is None else name
value = Parameter(name=name, value=value, lower=lower, upper=upper, step_size=step_size, label=label)
return value
@z.function(wraps="params", keepalive=True)
def assign_values_jit(
params: Parameter | Iterable[Parameter],
values: ztyping.NumericalScalarType | Iterable[ztyping.NumericalScalarType],
use_locking=False,
):
"""Assign values to parameters jitted.
This method can be significantly faster than `set_values`, however it expects the correct data-type and
cannot, for example, take a `FitResult` as input or function as a context manager. Only use when
performance is critical (such as inside a minimizer).
Args:
params: The parameters to assign the values to.
values: Values to assign to the parameters.
use_locking:
"""
for i, param in enumerate(params):
value = values[i]
if value.dtype != param.dtype:
value = znp.cast(value, param.dtype)
param.assign(value, read_value=False, use_locking=use_locking)
def assign_values(
params: Parameter | Iterable[Parameter],
values: ztyping.NumericalScalarType | Iterable[ztyping.NumericalScalarType],
use_locking=False,
allow_partial: bool | None = None,
):
"""Set the values of multiple parameters in a fast way.
In general, :meth:`set_values` is to be preferred. `assign_values` will ignore out-of-bounds errors,
does not offer a context-manager but is in general (an order of magnitude) faster.
Args:
params: Parameters to set the values.
values: List-like object that supports indexing.
use_locking: if true, lock the parameter to avoid race conditions.
allow_partial: Allow to set only parts of the parameters in case values is a `ZfitResult`
and not all are present in the
*values*. If False, *params* not in *values* will raise an error.
Note that setting this to true will also go with an empty values container.
Raises:
ValueError: If not all *params* are in *values* if *values* is a `FitResult` and `allow_partial` is `False`.
"""
if allow_partial is None:
allow_partial = False
params, values, _ = check_convert_param_values_assign(params, values, allow_partial=allow_partial)
assign_values_jit(params=params, values=values, use_locking=use_locking)
[docs]
def set_values(
params: Parameter | Iterable[Parameter],
values: (
ztyping.NumericalScalarType
| Iterable[ztyping.NumericalScalarType]
| ZfitResult
| Mapping[str | ZfitParameter, ztyping.NumericalScalarType]
| None,
) = None,
allow_partial: bool | None = None,
):
"""Set the values (using a context manager or not) of multiple parameters.
Args:
params: Parameters to set the values.
values: List-like object that supports indexing, a `FitResult` or a `Mapping` of parameters or parameter names to values.
allow_partial: Allow to set only parts of the parameters in case values is a `ZfitResult`
and not all are present in the
*values*. If False, *params* not in *values* will raise an error.
Note that setting this to true will also go with an empty values container.
Returns:
An object for a context manager (but can also be used without), can be ignored.
Raises:
ValueError: If the value is not between the limits of the parameter.
ValueError: If not all *params* are in *values* if *values* is a `FitResult` and `allow_partial` is `False`.
"""
if allow_partial is None:
allow_partial = False
params, values, _ = check_convert_param_values_assign(params, values, allow_partial)
def setter(values):
for i, param in enumerate(params):
param.set_value(values[i])
def getter():
return [param.value() for param in params]
return TemporarilySet(values, setter=setter, getter=getter)
def check_convert_param_values_assign(params, values, allow_partial=False):
"""Check if params are valid and convert them if necessary to be used with assign_values.
Args:
params: Parameters to set the values.
values: List-like object that supports indexing or a `ZfitResult` or a `Mapping`.
allow_partial: Allow to set only parts of the parameters in case values is a `ZfitResult` or a `Mapping`. Other parameters will not be updated.
More values than parameters, if they are given in a `Mapping` or `ZfitResult`, are allowed in any case.
Returns:
A tuple of (params, values)
"""
if isinstance(params, ZfitResult) and values is None:
params, values = None, params
elif isinstance(params, Mapping):
if values is not None:
msg = "Cannot set values to parameters if params are a Mapping."
raise ValueError(msg)
params, values = tuple(params.keys()), tuple(params.values())
elif not isinstance(values, ZfitResult):
params = convert_to_container(params)
if params is None:
msg = "No parameters given to set values to (values={values})."
raise ValueError(msg)
noparams = len(params) == 0
if noparams:
return params, values, True
if isinstance(values, ZfitResult):
result = values
new_params = []
values = []
if params is None:
params = result.params.keys()
for param in params:
if param in result.params:
values.append(result.params[param]["value"])
new_params.append(param)
elif not allow_partial:
msg = (
f"Cannot set {param} with {result} as it is not contained. To partially set"
f" the parameters (only those in the result), use allow_partial"
)
raise ValueError(msg)
params = new_params
elif isinstance(values, Mapping):
new_params = []
new_values = []
for param in params:
if param in values:
new_values.append(values[param])
new_params.append(param)
elif param.name in values:
new_values.append(values[param.name])
new_params.append(param)
elif not allow_partial:
msg = (
f"Cannot set {param} with {values!r} as it is not contained. To partially set"
f" the parameters (only those in the result), use allow_partial"
)
raise ValueError(msg)
values = new_values
params = new_params
else:
if not (tf.is_tensor(values) or isinstance(values, np.ndarray)):
values = convert_to_container(values)
lenvalues = len(values)
else:
values = znp.asarray(values, dtype=znp.float64)
values = znp.atleast_1d(values)
shape = values.shape
lenvalues = shape[0] if shape is not None else None
if lenvalues is not None and lenvalues != len(params):
msg = f"Incompatible length of parameters and values: {params}, {values}"
raise ValueError(msg)
not_param = [param for param in params if not isinstance(param, ZfitParameter)]
if not_param:
msg = f"The following are not parameters (but should be): {not_param}"
raise TypeError(msg)
non_independent_params = [param for param in params if not param.independent]
if non_independent_params:
msg = f"trying to set value of parameters that are not independent " f"{non_independent_params}"
raise ParameterNotIndependentError(msg)
values = znp.asarray(values, dtype=znp.float64)
return params, values, False # if it's empty
