from copy import deepcopy
from typing import Sequence, Tuple
import numpy as np
from ..C import CHI2, FVAL, GRAD, HESS, HESSP, RDATAS, RES, SCHI2, SRES
from .base import ObjectiveBase, ResultDict
[docs]class AggregatedObjective(ObjectiveBase):
"""Aggregates multiple objectives into one objective."""
[docs] def __init__(
self,
objectives: Sequence[ObjectiveBase],
x_names: Sequence[str] = None,
):
"""
Initialize objective.
Parameters
----------
objectives:
Sequence of pypesto.ObjectiveBase instances
x_names:
Sequence of names of the (optimized) parameters.
(Details see documentation of x_names in
:class:`pypesto.ObjectiveBase`)
"""
# input typechecks
if not isinstance(objectives, Sequence):
raise TypeError(
f'Objectives must be a Sequence, ' f'was {type(objectives)}.'
)
if not all(
isinstance(objective, ObjectiveBase) for objective in objectives
):
raise TypeError(
'Objectives must only contain elements of type'
'pypesto.Objective'
)
if not objectives:
raise ValueError('Length of objectives must be at least one')
self._objectives = objectives
super().__init__(x_names=x_names)
def __deepcopy__(self, memodict=None):
"""Create copy of objective."""
other = AggregatedObjective(
objectives=[deepcopy(objective) for objective in self._objectives],
x_names=deepcopy(self.x_names),
)
for key in set(self.__dict__.keys()) - {'_objectives', 'x_names'}:
other.__dict__[key] = deepcopy(self.__dict__[key])
return other
[docs] def check_mode(self, mode: str) -> bool:
"""See `ObjectiveBase` documentation."""
return all(
objective.check_mode(mode) for objective in self._objectives
)
[docs] def check_sensi_orders(
self,
sensi_orders: Tuple[int, ...],
mode: str,
) -> bool:
"""See `ObjectiveBase` documentation."""
return all(
objective.check_sensi_orders(sensi_orders, mode)
for objective in self._objectives
)
[docs] def call_unprocessed(
self,
x: np.ndarray,
sensi_orders: Tuple[int, ...],
mode: str,
**kwargs,
) -> ResultDict:
"""
See `ObjectiveBase` for more documentation.
Main method to overwrite from the base class. It handles and
delegates the actual objective evaluation.
"""
return aggregate_results(
[
objective.call_unprocessed(x, sensi_orders, mode, **kwargs)
for objective in self._objectives
]
)
[docs] def initialize(self):
"""See `ObjectiveBase` documentation."""
for objective in self._objectives:
objective.initialize()
[docs] def get_config(self) -> dict:
"""Return basic information of the objective configuration."""
info = super().get_config()
for n_obj, obj in enumerate(self._objectives):
info[f'objective_{n_obj}'] = obj.get_config()
return info
def aggregate_results(rvals: Sequence[ResultDict]) -> ResultDict:
"""
Aggregrate the results from the provided ResultDicts into a single one.
Parameters
----------
rvals:
results to aggregate
"""
# rvals are guaranteed to be consistent as _check_sensi_orders checks
# whether each objective can be called with the respective
# sensi_orders/mode
# sum over fval/grad/hess
result = {
key: sum(rval[key] for rval in rvals)
for key in [FVAL, CHI2, SCHI2, GRAD, HESS, HESSP]
if rvals[0].get(key, None) is not None
}
# extract rdatas and flatten
result[RDATAS] = []
for rval in rvals:
if RDATAS in rval:
result[RDATAS].extend(rval[RDATAS])
# initialize res and sres
if RES in rvals[0]:
res = np.asarray(rvals[0][RES])
else:
res = None
if SRES in rvals[0]:
sres = np.asarray(rvals[0][SRES])
else:
sres = None
# skip iobj=0 after initialization, stack matrices
for rval in rvals[1:]:
if res is not None:
res = np.hstack([res, np.asarray(rval[RES])])
if sres is not None:
sres = np.vstack([sres, np.asarray(rval[SRES])])
# fill res, sres into result
if res is not None:
result[RES] = res
if sres is not None:
result[SRES] = sres
return result