# Copyright (c) 2024 zfit
from __future__ import annotations
from functools import partial
from typing import Literal, Iterable
from typing import TYPE_CHECKING, List, Optional, Union
import pydantic
from pydantic import Field
from .serialmixin import SerializableMixin
from ..serialization.serializer import BaseRepr, Serializer
if TYPE_CHECKING:
import zfit
from collections.abc import Mapping
from collections.abc import Callable
from collections.abc import Iterable
import abc
import warnings
import tensorflow as tf
from ordered_set import OrderedSet
import zfit.z.numpy as znp
from .. import settings, z
from .interfaces import ZfitBinnedData, ZfitParameter, ZfitPDF, ZfitData
znp = z.numpy
from ..util import ztyping
from ..util.checks import NONE
from ..util.container import convert_to_container, is_container
from ..util.deprecation import deprecated_args
from ..util.exception import (
BehaviorUnderDiscussion,
BreakingAPIChangeError,
IntentionAmbiguousError,
NotExtendedPDFError,
)
from ..util.warnings import warn_advanced_feature
from ..z.math import (
autodiff_gradient,
autodiff_value_gradients,
automatic_value_gradients_hessian,
numerical_gradient,
numerical_value_gradient,
numerical_value_gradients_hessian,
)
from .baseobject import BaseNumeric, extract_filter_params
from .constraint import BaseConstraint
from .dependents import _extract_dependencies
from .interfaces import ZfitData, ZfitLoss, ZfitPDF, ZfitSpace
from .parameter import convert_to_parameters, set_values
DEFAULT_FULL_ARG = True
@z.function(wraps="loss")
def _unbinned_nll_tf(
model: ztyping.PDFInputType,
data: ztyping.DataInputType,
fit_range: ZfitSpace,
log_offset,
):
"""Return the unbinned negative log likelihood for a PDF.
Args:
model: |@doc:loss.init.model| PDFs that return the normalized probability for
*data* under the given parameters.
If multiple model and data are given, they will be used
in the same order to do a simultaneous fit. |@docend:loss.init.model|
data: |@doc:loss.init.data| Dataset that will be given to the *model*.
If multiple model and data are given, they will be used
in the same order to do a simultaneous fit. |@docend:loss.init.data|
fit_range:
Returns:
The unbinned nll value as a scalar
"""
if is_container(model):
nlls = [
_unbinned_nll_tf(model=p, data=d, fit_range=r, log_offset=log_offset)
for p, d, r in zip(model, data, fit_range)
]
nlls_summed = znp.sum(nlls, axis=0)
nll_finished = nlls_summed
else:
if fit_range is not None:
with data.set_data_range(fit_range):
probs = model.pdf(data, norm_range=fit_range)
else:
probs = model.pdf(data)
log_probs = znp.log(
probs + znp.asarray(1e-307, dtype=znp.float64)
) # minor offset to avoid NaNs from log(0)
if log_offset is None:
log_offset = znp.array([0.0], dtype=znp.float64)
nll = _nll_calc_unbinned_tf(
log_probs=log_probs,
weights=data.weights if data.weights is not None else None,
log_offset=log_offset,
)
nll_finished = nll
return nll_finished
@z.function(wraps="tensor", keepalive=True)
def _nll_calc_unbinned_tf(log_probs, weights, log_offset):
"""Calculate the negative log likelihood from the log probabilities."""
if weights is not None:
log_probs *= weights # because it's prob ** weights
if log_offset is not False:
log_probs -= log_offset
nll = -znp.sum(log_probs, axis=0)
# nll = -tfp.math.reduce_kahan_sum(input_tensor=log_probs, axis=0)
return nll
def _constraint_check_convert(constraints):
checked_constraints = []
for constr in constraints:
if isinstance(constr, BaseConstraint):
checked_constraints.append(constr)
else:
raise BreakingAPIChangeError(
"Constraints have to be of type `Constraint`, a simple"
" constraint from a function can be constructed with"
" `SimpleConstraint`."
)
return checked_constraints
class BaseLossRepr(BaseRepr):
_implementation = None
_owndict = pydantic.PrivateAttr(default_factory=dict)
hs3_type: Literal["BaseLoss"] = Field("BaseLoss", alias="type")
model: Union[
Serializer.types.PDFTypeDiscriminated,
List[Serializer.types.PDFTypeDiscriminated],
]
data: Union[
Serializer.types.DataTypeDiscriminated,
List[Serializer.types.DataTypeDiscriminated],
]
constraints: Optional[List[Serializer.types.ConstraintTypeDiscriminated]] = Field(
default_factory=list
)
options: Optional[Mapping] = Field(default_factory=dict)
@pydantic.validator("model", "data", "constraints", pre=True)
def _check_container(cls, v):
if cls.orm_mode(v):
v = convert_to_container(v, list)
return v
[docs]
class BaseLoss(ZfitLoss, BaseNumeric):
def __init__(
self,
model: ztyping.ModelsInputType,
data: ztyping.DataInputType,
fit_range: ztyping.LimitsTypeInput = None,
constraints: ztyping.ConstraintsTypeInput = None,
options: Mapping | None = None,
):
# first doc line left blank on purpose, subclass adds class docstring (Sphinx autodoc adds the two)
"""A "simultaneous fit" can be performed by giving one or more ``model``, ``data``, ``fit_range`` to the loss.
The length of each has to match the length of the others.
Args:
model: The model or models to evaluate the data on
data: Data to use
fit_range: The fitting range. It's the norm_range for the models (if
they
have a norm_range) and the data_range for the data.
constraints: A Tensor representing a loss constraint. Using
``zfit.constraint.*`` allows for easy use of predefined constraints.
options: Different options for the loss calculation.
"""
super().__init__(name=type(self).__name__, params={})
if fit_range is not None and all(fr is not None for fr in fit_range):
warnings.warn(
"The fit_range argument is depreceated and will maybe removed in future releases. "
"It is preferred to define the range in the space"
" when creating the data and the model.",
stacklevel=2,
)
model, data, fit_range = self._input_check(
pdf=model, data=data, fit_range=fit_range
)
self._model = model
self._data = data
self._fit_range = fit_range
options = self._check_init_options(options, data)
self._options = options
self._subtractions = {}
if constraints is None:
constraints = []
self._constraints = _constraint_check_convert(
convert_to_container(constraints, list)
)
self._is_precompiled = False
def _check_init_options(self, options, data):
try:
nevents = sum(d.nevents for d in data)
except (
RuntimeError
): # can happen if not yet sampled. What to do? Approx_nevents?
nevents = 150_000 # sensible default
options = {} if options is None else options
if options.get("numhess") is None:
options["numhess"] = True
if options.get("numgrad") is None:
options["numgrad"] = settings.options["numerical_grad"]
if options.get("kahansum") is None:
options["kahansum"] = (
nevents > 500_000
) # start using kahan if we have more than 500k events
if options.get("subtr_const") is None: # TODO: balance better?
# if nevents < 200_000:
# subtr_const = True
# elif nevents < 1_000_000:
# subtr_const = 'kahan'
# else:
# subtr_const = 'elewise'
subtr_const = True
options["subtr_const"] = subtr_const
return options
def __init_subclass__(cls, **kwargs):
super().__init_subclass__(**kwargs)
cls._name = "UnnamedSubBaseLoss"
def _get_params(
self,
floating: bool | None = True,
is_yield: bool | None = None,
extract_independent: bool | None = True,
) -> set[ZfitParameter]:
params = OrderedSet()
params = params.union(
*(
model.get_params(
floating=floating,
is_yield=is_yield,
extract_independent=extract_independent,
)
for model in self.model
)
)
params = params.union(
*(
constraint.get_params(
floating=floating,
is_yield=False,
extract_independent=extract_independent,
)
for constraint in self.constraints
)
)
return params
def _input_check(self, pdf, data, fit_range):
if isinstance(pdf, tuple):
raise TypeError("`pdf` has to be a pdf or a list of pdfs, not a tuple.")
if isinstance(data, tuple):
raise TypeError("`data` has to be a data or a list of data, not a tuple.")
# pdf, data = (convert_to_container(obj, non_containers=[tuple]) for obj in (pdf, data))
pdf, data = self._check_convert_model_data(pdf, data, fit_range)
# TODO: data, range consistency?
if fit_range is None:
fit_range = []
non_consistent = {"data": [], "model": [], "range": []}
for p, d in zip(pdf, data):
if p.norm != d.data_range:
non_consistent["data"].append(d)
non_consistent["model"].append(p)
non_consistent["range"].append((p.space, d.data_range))
fit_range.append(None)
if non_consistent["range"]: # TODO: test
warn_advanced_feature(
f"PDFs {non_consistent['model']} as "
f"well as `data` {non_consistent['data']}"
f" have different ranges {non_consistent['range']} they"
f" are defined in. The data range will cut the data while the"
f" norm range defines the normalization.",
identifier="inconsistent_fitrange",
)
else:
fit_range = convert_to_container(fit_range, non_containers=[tuple])
if not len(pdf) == len(data) == len(fit_range):
raise ValueError(
"pdf, data and fit_range don't have the same number of components:"
"\npdf: {}"
"\ndata: {}"
"\nfit_range: {}".format(pdf, data, fit_range)
)
# sanitize fit_range
fit_range = [
p._convert_sort_space(limits=range_) if range_ is not None else None
for p, range_ in zip(pdf, fit_range)
]
# TODO: sanitize pdf, data?
self.add_cache_deps(cache_deps=pdf)
self.add_cache_deps(cache_deps=data)
return pdf, data, fit_range
def _precompile(self):
if do_subtr := self._options.get("subtr_const", False):
if do_subtr is not True:
self._options["subtr_const_value"] = do_subtr
log_offset = self._options.get("subtr_const_value")
if log_offset is None:
from zfit import run
run.assert_executing_eagerly() # first time subtr
nevents_tot = znp.sum([d._approx_nevents for d in self.data])
log_offset_sum = (
self._call_value(
data=self.data,
model=self.model,
fit_range=self.fit_range,
constraints=self.constraints,
# presumably were not at the minimum,
# so the loss will decrease
log_offset=z.convert_to_tensor(0.0),
)
- 10000.0
)
log_offset = tf.stop_gradient(-znp.divide(log_offset_sum, nevents_tot))
self._options["subtr_const_value"] = log_offset
def _check_convert_model_data(self, model, data, fit_range):
model, data = tuple(convert_to_container(obj) for obj in (model, data))
model_checked = []
data_checked = []
for mod, dat in zip(model, data):
if not isinstance(dat, (ZfitData, ZfitBinnedData)):
if fit_range is not None:
raise TypeError(
"Fit range should not be used if data is not ZfitData."
)
if not isinstance(dat, (tf.Tensor, tf.Variable)):
try:
dat = z.convert_to_tensor(value=dat)
except TypeError:
raise TypeError(
f"Wrong type of dat ({type(dat)}). Has to be a `ZfitData` or convertible to a tf.Tensor"
)
# check dimension
from zfit import Data
dat = Data.from_tensor(obs=mod.space, tensor=dat)
model_checked.append(mod)
data_checked.append(dat)
return model_checked, data_checked
def _input_check_params(self, params):
if params is None:
params = list(self.get_params())
else:
params = convert_to_container(params)
return params
def add_constraints(self, constraints):
constraints = convert_to_container(constraints)
return self._add_constraints(constraints)
def _add_constraints(self, constraints):
constraints = _constraint_check_convert(
convert_to_container(constraints, container=list)
)
self._constraints.extend(constraints)
return constraints
@property
def name(self):
return self._name
@property
def model(self):
return self._model
@property
def data(self):
return self._data
@property
def fit_range(self):
return self._fit_range
@property
def constraints(self):
return self._constraints
def _get_dependencies(self): # TODO: fix, add constraints
pdf_dependents = _extract_dependencies(self.model)
pdf_dependents |= _extract_dependencies(self.constraints)
return pdf_dependents
@abc.abstractmethod
def _loss_func(self, model, data, fit_range, constraints, log_offset):
raise NotImplementedError
@property
def errordef(self) -> float | int:
return self._errordef
[docs]
def __call__(
self,
_x: ztyping.DataInputType = None,
# *, full: bool = None, # Not added, breaks iminuit.
) -> znp.array:
"""Calculate the loss value with the given input for the free parameters.
Args:
*positional*: Array-like argument to set the parameters. The order of the values correspond to
the position of the parameters in :py:meth:`~BaseLoss.get_params()` (called without any arguments).
For more detailed control, it is always possible to wrap :py:meth:`~BaseLoss.value()` and set the
desired parameters manually.
full: |@doc:loss.value.full| If True, return the full loss value, otherwise
allow for the removal of constants and only return
the part that depends on the parameters. Constants
don't matter for the task of optimization, but
they can greatly help with the numerical stability of the loss function. |@docend:loss.value.full|
Returns:
Calculated loss value as a scalar.
"""
if _x is None:
raise BehaviorUnderDiscussion(
"Currently, calling a loss requires to give the arguments explicitly."
" If you think this behavior should be changed, please open an issue"
" https://github.com/zfit/zfit/issues/new/choose"
)
if isinstance(_x, dict):
raise TypeError(
"Dicts are not supported when calling a loss, only array-like values."
)
if _x is None:
return self.value(full=True) # has to be full, otherwise iminuit breaks
else:
params = self.get_params()
with set_values(params, _x):
return self.value(full=True)
[docs]
def value(self, *, full: bool = None) -> znp.ndarray:
"""Calculate the loss value with the current values of the free parameters.
Args:
full: |@doc:loss.value.full| If True, return the full loss value, otherwise
allow for the removal of constants and only return
the part that depends on the parameters. Constants
don't matter for the task of optimization, but
they can greatly help with the numerical stability of the loss function. |@docend:loss.value.full|
Returns:
Calculated loss value as a scalar.
"""
if not self._is_precompiled:
self._precompile()
self._is_precompiled = True
if full is None:
full = DEFAULT_FULL_ARG
if full:
log_offset = 0.0
else:
log_offset = self._options.get("subtr_const_value")
if log_offset is not None:
log_offset = z.convert_to_tensor(log_offset)
# log_offset = z.convert_to_tensor(log_offset)
value = self._call_value(
self.model, self.data, self.fit_range, self.constraints, log_offset
)
return value
def _call_value(self, model, data, fit_range, constraints, log_offset):
value = self._value(
model=model,
data=data,
fit_range=fit_range,
constraints=constraints,
log_offset=log_offset,
)
# if self._subtractions.get('kahan') is None:
# self._subtractions['kahan'] = value
# value_subtracted = (value[0] - self._subtractions['kahan'][0]) - (
# value[1] - self._subtractions['kahan'][1])
# return value_subtracted
return value
# value = value_substracted[0] - value_substracted[1]
def _value(self, model, data, fit_range, constraints, log_offset):
return self._loss_func(
model=model,
data=data,
fit_range=fit_range,
constraints=constraints,
log_offset=log_offset,
)
def __add__(self, other):
if not isinstance(other, BaseLoss):
raise TypeError(
"Has to be a subclass of `BaseLoss` or overwrite `__add__`."
)
if type(other) != type(self):
raise ValueError("cannot safely add two different kind of loss.")
model = self.model + other.model
data = self.data + other.data
fit_range = self.fit_range + other.fit_range
constraints = self.constraints + other.constraints
kwargs = dict(model=model, data=data, constraints=constraints)
if any(fitrng is not None for fitrng in fit_range):
kwargs["fit_range"] = fit_range
return type(self)(**kwargs)
[docs]
def gradient(
self, params: ztyping.ParamTypeInput = None, *, numgrad=None
) -> list[tf.Tensor]:
"""Calculate the gradient of the loss with respect to the given parameters.
Args:
params: The parameters with respect to which the gradient is calculated. If `None`, all parameters
are used.
numgrad: |@doc:loss.args.numgrad| If ``True``, calculate the numerical gradient/Hessian
instead of using the automatic one. This is
usually slower if called repeatedly but can
be used if the automatic gradient fails (e.g. if
the model is not differentiable, written not in znp.* etc).
Default will fall back to what the loss is set to. |@docend:loss.args.numgrad|
Returns:
The gradient of the loss with respect to the given parameters.
"""
params = self._input_check_params(params)
numgrad = self._options["numgrad"] if numgrad is None else numgrad
params = {p.name: p for p in params}
if not self._is_precompiled:
self._precompile()
self._is_precompiled = True
return self._gradient(params=params, numgrad=numgrad)
def gradients(self, *args, **kwargs):
raise BreakingAPIChangeError(
"`gradients` is deprecated, use `gradient` instead."
)
@z.function(wraps="loss")
def _gradient(self, params, numgrad):
params = tuple(params.values())
self_value = partial(self.value, full=False)
if numgrad:
gradient = numerical_gradient(self_value, params=params)
else:
gradient = autodiff_gradient(self_value, params=params)
return gradient
[docs]
def value_gradient(
self,
params: ztyping.ParamTypeInput = None,
*,
full: bool = None,
numgrad: bool = None,
) -> tuple[tf.Tensor, tf.Tensor]:
"""Calculate the loss value and the gradient with the current values of the free parameters.
Args:
params: The parameters to calculate the gradient for. If not given, all free parameters are used.
full: |@doc:loss.value.full| If True, return the full loss value, otherwise
allow for the removal of constants and only return
the part that depends on the parameters. Constants
don't matter for the task of optimization, but
they can greatly help with the numerical stability of the loss function. |@docend:loss.value.full|
numgrad: |@doc:loss.args.numgrad| If ``True``, calculate the numerical gradient/Hessian
instead of using the automatic one. This is
usually slower if called repeatedly but can
be used if the automatic gradient fails (e.g. if
the model is not differentiable, written not in znp.* etc).
Default will fall back to what the loss is set to. |@docend:loss.args.numgrad|
Returns:
Calculated loss value as a scalar and the gradient as a tensor.
"""
params = self._input_check_params(params)
numgrad = self._options["numgrad"]
params = {p.name: p for p in params}
if full is None:
full = DEFAULT_FULL_ARG
if not self._is_precompiled:
self._precompile()
self._is_precompiled = True
return self._value_gradient(params=params, numgrad=numgrad, full=full)
def value_gradients(self, *args, **kwargs):
raise BreakingAPIChangeError(
"`value_gradients` is deprecated, use `value_gradient` instead."
)
@z.function(wraps="loss")
def _value_gradient(self, params, numgrad=False, *, full: bool = None):
params = tuple(params.values())
if full is None:
full = DEFAULT_FULL_ARG
self_value = partial(self.value, full=full)
if numgrad:
value, gradient = numerical_value_gradient(self_value, params=params)
else:
value, gradient = autodiff_value_gradients(self_value, params=params)
return value, gradient
[docs]
def hessian(
self,
params: ztyping.ParamTypeInput = None,
hessian=None,
*,
numgrad: bool = None,
):
"""Calculate the hessian of the loss with respect to the given parameters.
Args:
params: The parameters with respect to which the hessian is calculated. If `None`, all parameters
are used.
hessian: Can be 'full' or 'diag'.
numgrad: |@doc:loss.args.numgrad| If ``True``, calculate the numerical gradient/Hessian
instead of using the automatic one. This is
usually slower if called repeatedly but can
be used if the automatic gradient fails (e.g. if
the model is not differentiable, written not in znp.* etc).
Default will fall back to what the loss is set to. |@docend:loss.args.numgrad|
"""
params = self._input_check_params(params)
if not self._is_precompiled:
self._precompile()
self._is_precompiled = True
return self.value_gradient_hessian(params=params, hessian=hessian, full=False)[
2
]
[docs]
def value_gradient_hessian(
self,
params: ztyping.ParamTypeInput = None,
hessian=None,
*,
full: bool = None,
numgrad=None,
) -> tuple[tf.Tensor, tf.Tensor, tf.Tensor]:
"""Calculate the loss value, the gradient and the hessian with the current values of the free parameters.
Args:
params: The parameters to calculate the gradient for. If not given, all free parameters are used.
hessian: Can be 'full' or 'diag'.
full: |@doc:loss.value.full| If True, return the full loss value, otherwise
allow for the removal of constants and only return
the part that depends on the parameters. Constants
don't matter for the task of optimization, but
they can greatly help with the numerical stability of the loss function. |@docend:loss.value.full|
numgrad: |@doc:loss.args.numgrad| If ``True``, calculate the numerical gradient/Hessian
instead of using the automatic one. This is
usually slower if called repeatedly but can
be used if the automatic gradient fails (e.g. if
the model is not differentiable, written not in znp.* etc).
Default will fall back to what the loss is set to. |@docend:loss.args.numgrad|
Returns:
Calculated loss value as a scalar, the gradient as a tensor and the hessian as a tensor.
"""
params = self._input_check_params(params)
numgrad = self._options["numhess"] if numgrad is None else numgrad
params = {p.name: p for p in params}
if full is None:
full = DEFAULT_FULL_ARG
if not self._is_precompiled:
self._precompile()
self._is_precompiled = True
vals = self._value_gradient_hessian(
params=params, hessian=hessian, numerical=numgrad, full=full
)
vals = vals[0], z.convert_to_tensor(vals[1]), vals[2]
return vals
def value_gradients_hessian(self, *args, **kwargs):
raise BreakingAPIChangeError(
"`value_gradients_hessian` is deprecated, use `value_gradient_hessian` instead."
)
@z.function(wraps="loss")
def _value_gradient_hessian(
self, params, hessian, numerical=False, *, full: bool = None
):
params = tuple(params.values())
self_value = partial(self.value, full=full)
if numerical:
return numerical_value_gradients_hessian(
func=self_value, gradient=self.gradient, params=params, hessian=hessian
)
else:
return automatic_value_gradients_hessian(
self_value, params=params, hessian=hessian
)
def __repr__(self) -> str:
class_name = repr(self.__class__)[:-2].split(".")[-1]
return (
f"<{class_name} "
f"model={one_two_many([model.name for model in self.model])} "
f"data={one_two_many([data.name for data in self.data])} "
f'constraints={one_two_many(self.constraints, many="True")} '
f">"
)
def __str__(self) -> str:
class_name = repr(self.__class__)[:-2].split(".")[-1]
return (
f"<{class_name}"
f" model={one_two_many([model for model in self.model])}"
f" data={one_two_many([data for data in self.data])}"
f' constraints={one_two_many(self.constraints, many="True")}'
f">"
)
def one_two_many(values, n=3, many="multiple"):
values = convert_to_container(values)
if len(values) > n:
values = many
return values
class BaseUnbinnedNLL(BaseLoss, SerializableMixin):
def create_new(
self,
model: ZfitPDF | Iterable[ZfitPDF] | None = NONE,
data: ZfitData | Iterable[ZfitData] | None = NONE,
fit_range=NONE,
constraints=NONE,
options=NONE,
):
r"""Create a new loss from the current loss and replacing what is given as the arguments.
This creates a "copy" of the current loss but replaces any argument that is explicitly given.
Equivalent to creating a new instance but with some arguments taken.
A loss has more than a model and data (and constraints), it can have internal optimizations
and more that may do alter the behavior of a naive re-instantiation in unpredictable ways.
Args:
model: If not given, the current one will be used.
|@doc:loss.init.model| PDFs that return the normalized probability for
*data* under the given parameters.
If multiple model and data are given, they will be used
in the same order to do a simultaneous fit. |@docend:loss.init.model|
data: If not given, the current one will be used.
|@doc:loss.init.data| Dataset that will be given to the *model*.
If multiple model and data are given, they will be used
in the same order to do a simultaneous fit. |@docend:loss.init.data|
fit_range:
constraints: If not given, the current one will be used.
|@doc:loss.init.constraints| Auxiliary measurements ("constraints")
that add a likelihood term to the loss.
.. math::
\mathcal{L}(\theta) = \mathcal{L}_{unconstrained} \prod_{i} f_{constr_i}(\theta)
Usually, an auxiliary measurement -- by its very nature -S should only be added once
to the loss. zfit does not automatically deduplicate constraints if they are given
multiple times, leaving the freedom for arbitrary constructs.
Constraints can also be used to restrict the loss by adding any kinds of penalties. |@docend:loss.init.constraints|
options: If not given, the current one will be used.
|@doc:loss.init.options| Additional options (as a dict) for the loss.
Current possibilities include:
- 'subtr_const' (default True): subtract from each points
log probability density a constant that
is approximately equal to the average log probability
density in the very first evaluation before
the summation. This brings the initial loss value closer to 0 and increases,
especially for large datasets, the numerical stability.
The value will be stored ith 'subtr_const_value' and can also be given
directly.
The subtraction should not affect the minimum as the absolute
value of the NLL is meaningless. However,
with this switch on, one cannot directly compare
different likelihoods absolute value as the constant
may differ! Use ``create_new`` in order to have a comparable likelihood
between different losses or use the ``full`` argument in the value function
to calculate the full loss with all constants.
These settings may extend over time. In order to make sure that a loss is the
same under the same data, make sure to use ``create_new`` instead of instantiating
a new loss as the former will automatically overtake any relevant constants
and behavior. |@docend:loss.init.options|
"""
if model is NONE:
model = self.model
if data is NONE:
data = self.data
if fit_range is NONE:
fit_range = self.fit_range
if constraints is NONE:
constraints = self.constraints
if constraints is not None:
constraints = constraints.copy()
if options is NONE:
options = self._options
if isinstance(options, dict):
options = options.copy()
return type(self)(
model=model,
data=data,
fit_range=fit_range,
constraints=constraints,
options=options,
)
[docs]
class UnbinnedNLL(BaseUnbinnedNLL):
_name = "UnbinnedNLL"
def __init__(
self,
model: ZfitPDF | Iterable[ZfitPDF],
data: ZfitData | Iterable[ZfitData],
fit_range=None,
constraints: ztyping.ConstraintsInputType = None,
options: Mapping[str, object] | None = None,
):
r"""Unbinned Negative Log Likelihood.
|@doc:loss.init.explain.unbinnednll| The unbinned log likelihood can be written as
.. math::
\mathcal{L}_{non-extended}(x | \theta) = \prod_{i} f_{\theta} (x_i)
where :math:`x_i` is a single event from the dataset *data* and f is the *model*. |@docend:loss.init.explain.unbinnednll|
|@doc:loss.init.explain.simultaneous| A simultaneous fit can be performed by giving one or more ``model``, ``data``, to the loss. The
length of each has to match the length of the others
.. math::
\mathcal{L}_{simultaneous}(\theta | {data_0, data_1, ..., data_n})
= \prod_{i} \mathcal{L}(\theta_i, data_i)
where :math:`\theta_i` is a set of parameters and
a subset of :math:`\theta` |@docend:loss.init.explain.simultaneous|
|@doc:loss.init.explain.negativelog| For optimization purposes, it is often easier
to minimize a function and to use a log transformation. The actual loss is given by
.. math::
\mathcal{L} = - \sum_{i}^{n} ln(f(\theta|x_i))
and therefore being called "negative log ..." |@docend:loss.init.explain.negativelog|
|@doc:loss.init.explain.weightednll| If the dataset has weights, a weighted likelihood will be constructed instead
.. math::
\mathcal{L} = - \sum_{i}^{n} w_i \cdot ln(f(\theta|x_i))
Note that this is not a real likelihood anymore! Calculating uncertainties
can be done with hesse (as it has a correction) but will yield wrong
results with profiling methods. The minimum is however fully valid. |@docend:loss.init.explain.weightednll|
Args:
model: |@doc:loss.init.model| PDFs that return the normalized probability for
*data* under the given parameters.
If multiple model and data are given, they will be used
in the same order to do a simultaneous fit. |@docend:loss.init.model|
data: |@doc:loss.init.data| Dataset that will be given to the *model*.
If multiple model and data are given, they will be used
in the same order to do a simultaneous fit. |@docend:loss.init.data|
constraints: |@doc:loss.init.constraints| Auxiliary measurements ("constraints")
that add a likelihood term to the loss.
.. math::
\mathcal{L}(\theta) = \mathcal{L}_{unconstrained} \prod_{i} f_{constr_i}(\theta)
Usually, an auxiliary measurement -- by its very nature -S should only be added once
to the loss. zfit does not automatically deduplicate constraints if they are given
multiple times, leaving the freedom for arbitrary constructs.
Constraints can also be used to restrict the loss by adding any kinds of penalties. |@docend:loss.init.constraints|
options: If not given, the current one will be used.
|@doc:loss.init.options| Additional options (as a dict) for the loss.
Current possibilities include:
- 'subtr_const' (default True): subtract from each points
log probability density a constant that
is approximately equal to the average log probability
density in the very first evaluation before
the summation. This brings the initial loss value closer to 0 and increases,
especially for large datasets, the numerical stability.
The value will be stored ith 'subtr_const_value' and can also be given
directly.
The subtraction should not affect the minimum as the absolute
value of the NLL is meaningless. However,
with this switch on, one cannot directly compare
different likelihoods absolute value as the constant
may differ! Use ``create_new`` in order to have a comparable likelihood
between different losses or use the ``full`` argument in the value function
to calculate the full loss with all constants.
These settings may extend over time. In order to make sure that a loss is the
same under the same data, make sure to use ``create_new`` instead of instantiating
a new loss as the former will automatically overtake any relevant constants
and behavior. |@docend:loss.init.options|
"""
super().__init__(
model=model,
data=data,
fit_range=fit_range,
constraints=constraints,
options=options,
)
self._errordef = 0.5
extended_pdfs = [pdf for pdf in self.model if pdf.is_extended]
if extended_pdfs and type(self) == UnbinnedNLL:
warn_advanced_feature(
f"Extended PDFs ({extended_pdfs}) are given to a normal UnbinnedNLL. "
f" This won't take the yield "
"into account and simply treat the PDFs as non-extended PDFs. To create an "
"extended NLL, use the `ExtendedUnbinnedNLL`.",
identifier="extended_in_UnbinnedNLL",
)
def _loss_func(self, model, data, fit_range, constraints, log_offset):
return self._loss_func_watched(
data=data,
model=model,
fit_range=fit_range,
constraints=constraints,
log_offset=log_offset,
)
@property
def is_extended(self):
return False
@z.function(wraps="loss")
def _loss_func_watched(self, data, model, fit_range, constraints, log_offset):
nll = _unbinned_nll_tf(
model=model, data=data, fit_range=fit_range, log_offset=log_offset
)
if constraints:
constraints = z.reduce_sum([c.value() for c in constraints])
nll += constraints
return nll
def _get_params(
self,
floating: bool | None = True,
is_yield: bool | None = None,
extract_independent: bool | None = True,
) -> set[ZfitParameter]:
if not self.is_extended:
is_yield = False # the loss does not depend on the yields
return super()._get_params(floating, is_yield, extract_independent)
class UnbinnedNLLRepr(BaseLossRepr):
_implementation = UnbinnedNLL
hs3_type: Literal["UnbinnedNLL"] = pydantic.Field("UnbinnedNLL", alias="type")
[docs]
class ExtendedUnbinnedNLL(BaseUnbinnedNLL):
def __init__(
self,
model: ZfitPDF | Iterable[ZfitPDF],
data: ZfitData | Iterable[ZfitData],
fit_range=None,
constraints: ztyping.ConstraintsInputType = None,
options: Mapping[str, object] | None = None,
):
r"""An Unbinned Negative Log Likelihood with an additional poisson term for the number of events in the dataset.
|@doc:loss.init.explain.unbinnednll| The unbinned log likelihood can be written as
.. math::
\mathcal{L}_{non-extended}(x | \theta) = \prod_{i} f_{\theta} (x_i)
where :math:`x_i` is a single event from the dataset *data* and f is the *model*. |@docend:loss.init.explain.unbinnednll|
|@doc:loss.init.explain.extendedterm| The extended likelihood has an additional term
.. math::
\mathcal{L}_{extended term} = poiss(N_{tot}, N_{data})
= N_{data}^{N_{tot}} \frac{e^{- N_{data}}}{N_{tot}!}
and the extended likelihood is the product of both. |@docend:loss.init.explain.extendedterm|
|@doc:loss.init.explain.simultaneous| A simultaneous fit can be performed by giving one or more ``model``, ``data``, to the loss. The
length of each has to match the length of the others
.. math::
\mathcal{L}_{simultaneous}(\theta | {data_0, data_1, ..., data_n})
= \prod_{i} \mathcal{L}(\theta_i, data_i)
where :math:`\theta_i` is a set of parameters and
a subset of :math:`\theta` |@docend:loss.init.explain.simultaneous|
|@doc:loss.init.explain.negativelog| For optimization purposes, it is often easier
to minimize a function and to use a log transformation. The actual loss is given by
.. math::
\mathcal{L} = - \sum_{i}^{n} ln(f(\theta|x_i))
and therefore being called "negative log ..." |@docend:loss.init.explain.negativelog|
|@doc:loss.init.explain.weightednll| If the dataset has weights, a weighted likelihood will be constructed instead
.. math::
\mathcal{L} = - \sum_{i}^{n} w_i \cdot ln(f(\theta|x_i))
Note that this is not a real likelihood anymore! Calculating uncertainties
can be done with hesse (as it has a correction) but will yield wrong
results with profiling methods. The minimum is however fully valid. |@docend:loss.init.explain.weightednll|
"""
super().__init__(
model=model,
data=data,
constraints=constraints,
options=options,
fit_range=fit_range,
)
self._errordef = 0.5
@z.function(wraps="loss")
def _loss_func(self, model, data, fit_range, constraints, log_offset):
nll = _unbinned_nll_tf(
model=model, data=data, fit_range=fit_range, log_offset=log_offset
)
if constraints:
constraints = z.reduce_sum([c.value() for c in constraints])
nll += constraints
yields = []
nevents_collected = []
for mod, dat in zip(model, data):
if not mod.is_extended:
raise NotExtendedPDFError(
f"The pdf {mod} is not extended but has to be (for an extended fit)"
)
nevents = dat.n_events if dat.weights is None else z.reduce_sum(dat.weights)
nevents = tf.cast(nevents, tf.float64)
nevents_collected.append(nevents)
yields.append(mod.get_yield())
yields = znp.stack(yields, axis=0)
nevents_collected = znp.stack(nevents_collected, axis=0)
term_new = tf.nn.log_poisson_loss(
nevents_collected, znp.log(yields), compute_full_loss=log_offset is False
)
if log_offset is not False:
log_offset = znp.asarray(log_offset, dtype=znp.float64)
term_new += log_offset
nll += znp.sum(term_new, axis=0)
return nll
@property
def is_extended(self):
return True
def _get_params(
self,
floating: bool | None = True,
is_yield: bool | None = None,
extract_independent: bool | None = True,
) -> set[ZfitParameter]:
return super()._get_params(floating, is_yield, extract_independent)
class ExtendedUnbinnedNLLRepr(BaseLossRepr):
_implementation = ExtendedUnbinnedNLL
hs3_type: Literal["ExtendedUnbinnedNLL"] = pydantic.Field(
"ExtendedUnbinnedNLL", alias="type"
)
[docs]
class SimpleLoss(BaseLoss):
_name = "SimpleLoss"
@deprecated_args(None, "Use params instead.", ("deps", "dependents"))
def __init__(
self,
func: Callable,
params: Iterable[zfit.Parameter] = None,
errordef: float | None = None,
# legacy
deps: Iterable[zfit.Parameter] = NONE,
dependents: Iterable[zfit.Parameter] = NONE,
):
r"""Loss from a (function returning a) Tensor.
This allows for a very generic loss function as the functions only restriction is that is
should depend on ``zfit.Parameter``.
Args:
func: Callable that constructs the loss and returns a tensor without taking an argument.
params: The dependents (independent ``zfit.Parameter``) of the loss. Essentially the (free) parameters that
the ``func`` depends on.
errordef: Definition of which change in the loss corresponds to a change of 1 sigma.
For example, 1 for Chi squared, 0.5 for negative log-likelihood.
Usage:
.. code:: python
import zfit
import zfit.z.numpy as znp
import tensorflow as tf
param1 = zfit.Parameter('param1', 5, 1, 10)
# we can build a model here if we want, but in principle, it's not necessary
x = znp.random.uniform(size=(100,))
y = x * tf.random.normal(mean=4, stddev=0.1, shape=x.shape, dtype=znp.float64)
def squared_loss(params):
param1 = params[0]
y_pred = x*param1 # this is very simple, but we can of course use any
# zfit PDF or Func inside
squared = (y_pred - y)**2
mse = znp.mean(squared)
return mse
loss = zfit.loss.SimpleLoss(squared_loss, param1, errordef=1)
which can then be used in combination with any zfit minimizer such as Minuit
.. code:: python
minimizer = zfit.minimize.Minuit()
result = minimizer.minimize(loss)
"""
super().__init__(model=[], data=[], options={"subtr_const": False})
if dependents is not NONE and params is None:
params = dependents
elif deps is not NONE and params is None: # depreceation
params = deps
elif params is None: # legacy, remove in 0.7
raise BreakingAPIChangeError(
"params need to be specified explicitly due to the upgrade to 0.4."
"More information can be found in the upgrade guide on the website."
)
if hasattr(func, "errordef"):
if errordef is not None:
raise ValueError(
"errordef is not allowed if func has an errordef attribute or vice versa."
)
errordef = func.errordef
if errordef is None:
raise ValueError(
f"{self} cannot minimize {func} as `errordef` is missing: "
f"it has to be set as an attribute. Typically 1 (chi2) or 0.5 (NLL)."
)
self._simple_func = func
self._errordef = errordef
params = convert_to_parameters(params, prefer_constant=False)
self._params = params
self._simple_func_params = _extract_dependencies(params)
def _get_dependencies(self):
dependents = self._simple_func_params
return dependents
def _get_params(
self,
floating: bool | None = True,
is_yield: bool | None = None,
extract_independent: bool | None = True,
) -> set[ZfitParameter]:
params = super()._get_params(floating, is_yield, extract_independent)
own_params = extract_filter_params(
self._params, floating=floating, extract_independent=extract_independent
)
params = params.union(own_params)
return params
@property
def errordef(self):
errordef = self._errordef
if errordef is None:
raise RuntimeError("For this SimpleLoss, no error calculation is possible.")
else:
return errordef
# @z.function(wraps='loss')
def _loss_func(self, model, data, fit_range, constraints=None, log_offset=None):
if log_offset not in (None, False):
raise ValueError("log_offset is not allowed for a SimpleLoss")
try:
params = self._simple_func_params
params = tuple(params)
value = self._simple_func(params)
except TypeError as error:
if "takes 0 positional arguments but 1 was given" in str(error):
value = self._simple_func()
else:
raise error
return z.convert_to_tensor(value)
def __add__(self, other):
raise IntentionAmbiguousError(
"Cannot add a SimpleLoss, 'addition' of losses can mean anything."
"Add them manually"
)
def create_new(
self,
func: Callable = NONE,
params: Iterable[zfit.Parameter] = NONE,
errordef: float | None = NONE,
):
if func is NONE:
func = self._simple_func
if params is NONE:
params = self._params
if errordef is NONE:
errordef = self.errordef
return type(self)(func=func, params=params, errordef=errordef)
