Source code for orion.algo.base

# -*- coding: utf-8 -*-
"""
Base Search Algorithm
=====================

Formulation of a general search algorithm with respect to some objective.
Algorithm implementations must inherit from `orion.algo.base.OptimizationAlgorithm`.

"""
import copy
import hashlib
import logging
from abc import ABCMeta, abstractmethod

from orion.algo.space import Fidelity
from orion.core.utils import Factory

log = logging.getLogger(__name__)


def infer_trial_id(point):
    """Compute a hashing of a point"""
    return hashlib.md5(str(list(point)).encode("utf-8")).hexdigest()


# pylint: disable=too-many-public-methods
[docs]class BaseAlgorithm(object, metaclass=ABCMeta): """Base class describing what an algorithm can do. Parameters ---------- space : `orion.algo.space.Space` Definition of a problem's parameter space. kwargs : dict Tunable elements of a particular algorithm, a dictionary from hyperparameter names to values. Notes ----- We are using the No Free Lunch theorem's [1]_[3]_ formulation of an `BaseAlgorithm`. We treat it as a part of a procedure which in each iteration suggests a sample of the parameter space of the problem as a candidate solution and observes the results of its evaluation. **Developer Note**: Each algorithm's complete specification, i.e. implementation of its methods and parameters of its own, lies in a separate concrete algorithm class, which must be an **immediate** subclass of `BaseAlgorithm`. [The reason for this is current implementation of `orion.core.utils.Factory` metaclass which uses `BaseAlgorithm.__subclasses__()`.] Second, one must declare an algorithm's own parameters (tunable elements which could be set by configuration). This is done by passing them to `BaseAlgorithm.__init__()` by calling Python's super with a `Space` object as a positional argument plus algorithm's own parameters as keyword arguments. The keys of the keyword arguments passed to `BaseAlgorithm.__init__()` are interpreted as the algorithm's parameter names. So for example, a subclass could be as simple as this (regarding the logistics, not an actual algorithm's implementation): Examples -------- .. code-block:: python :linenos: :emphasize-lines: 7 from orion.algo.base import BaseAlgorithm from orion.algo.space import (Integer, Space) class MySimpleAlgo(BaseAlgorithm): def __init__(self, space, multiplier=1, another_param="a string param"): super().__init__(space, multiplier=multiplier, another_param=another_param) def suggest(self, num=1): print(self.another_param) return list(map(lambda x: tuple(map(lambda y: self.multiplier * y, x)), self.space.sample(num))) def observe(self, points, results): pass dim = Integer('named_param', 'norm', 3, 2, shape=(2, 3)) s = Space() s.register(dim) algo = MySimpleAlgo(s, 2, "I am just sampling!") algo.suggest() References ---------- .. [1] D. H. Wolpert and W. G. Macready, “No Free Lunch Theorems for Optimization,” IEEE Transactions on Evolutionary Computation, vol. 1, no. 1, pp. 67–82, Apr. 1997. .. [2] W. G. Macready and D. H. Wolpert, “What Makes An Optimization Problem Hard?,” Complexity, vol. 1, no. 5, pp. 40–46, 1996. .. [3] D. H. Wolpert and W. G. Macready, “No Free Lunch Theorems for Search,” Technical Report SFI-TR-95-02-010, Santa Fe Institute, 1995. """ requires_type = None requires_shape = None requires_dist = None def __init__(self, space, **kwargs): log.debug( "Creating Algorithm object of %s type with parameters:\n%s", type(self).__name__, kwargs, ) self._trials_info = {} # Stores Unique Trial -> Result self._space = space self._param_names = list(kwargs.keys()) # Instantiate tunable parameters of an algorithm for varname, param in kwargs.items(): # Check if tunable element is another algorithm if isinstance(param, dict) and len(param) == 1: subalgo_type = list(param)[0] subalgo_kwargs = param[subalgo_type] if isinstance(subalgo_kwargs, dict): param = OptimizationAlgorithm(subalgo_type, space, **subalgo_kwargs) elif ( isinstance(param, str) and param.lower() in OptimizationAlgorithm.types ): # pylint: disable=too-many-function-args param = OptimizationAlgorithm(param, space) elif varname == "seed": self.seed_rng(param) setattr(self, varname, param)
[docs] def seed_rng(self, seed): """Seed the state of the random number generator. :param seed: Integer seed for the random number generator. .. note:: This methods does nothing if the algorithm is deterministic. """ pass
@property def state_dict(self): """Return a state dict that can be used to reset the state of the algorithm.""" return {"_trials_info": copy.deepcopy(self._trials_info)}
[docs] def set_state(self, state_dict): """Reset the state of the algorithm based on the given state_dict :param state_dict: Dictionary representing state of an algorithm """ self._trials_info = state_dict.get("_trials_info")
[docs] def format_point(self, point): """Format point based on space transformations This will apply the reverse transformation on the point and then transform it again. Some transformations are lossy and thus the points suggested by the algorithm could be different when returned to `observe`. Using `format_point` makes it possible for the algorithm to see the final version of the point after back and forth transformations. This way it can recognise the point in `observe` and also avoid duplicates that would have gone unnoticed during suggestion. Parameters ---------- point : tuples of array-likes Points from a `orion.algo.space.Space`. """ point = tuple(point) if hasattr(self.space, "transform"): point = self.space.transform(self.space.reverse(point)) return point
[docs] def get_id(self, point, ignore_fidelity=False): """Compute a unique hash for a point based on params Parameters ---------- point : tuples of array-likes Points from a `orion.algo.space.Space`. ignore_fidelity: bool, optional If True, the fidelity dimension is ignored when computing a unique hash for the trial. Defaults to False. """ # Apply transforms and reverse to see data as it would come from DB # (Some transformations looses some info. ex: Precision transformation) point = list(self.format_point(point)) if ignore_fidelity: non_fidelity_dims = point[0 : self.fidelity_index] non_fidelity_dims.extend(point[self.fidelity_index + 1 :]) point = non_fidelity_dims return hashlib.md5(str(point).encode("utf-8")).hexdigest()
@property def fidelity_index(self): """Compute the index of the point where fidelity is. Returns None if there is no fidelity dimension. """ def _is_fidelity(dim): return dim.type == "fidelity" fidelity_index = [ i for i, dim in enumerate(self.space.values()) if _is_fidelity(dim) ] if fidelity_index: return fidelity_index[0] return None
[docs] @abstractmethod def suggest(self, num): """Suggest a `num` of new sets of parameters. Parameters ---------- num: int Number of points to suggest. The algorithm may return less than the number of points requested. Returns ------- list of points or None A list of lists representing points suggested by the algorithm. The algorithm may opt out if it cannot make a good suggestion at the moment (it may be waiting for other trials to complete), in which case it will return None. Notes ----- New parameters must be compliant with the problem's domain `orion.algo.space.Space`. """ pass
[docs] def observe(self, points, results): """Observe the `results` of the evaluation of the `points` in the process defined in user's script. Parameters ---------- points : list of tuples of array-likes Points from a `orion.algo.space.Space`. results : list of dicts Contains the result of an evaluation; partial information about the black-box function at each point in `params`. Result ------ objective : numeric Evaluation of this problem's objective function. gradient : 1D array-like, optional Contains values of the derivatives of the `objective` function with respect to `params`. constraint : list of numeric, optional List of constraints expression evaluation which must be greater or equal to zero by the problem's definition. """ for point, result in zip(points, results): if not self.has_observed(point): self.register(point, result)
[docs] def register(self, point, result=None): """Save the point as one suggested or observed by the algorithm Parameters ---------- point : array-likes Point from a `orion.algo.space.Space`. result : dict or None, optional The result of an evaluation; partial information about the black-box function at each point in `params`. None is suggested and not yet completed. """ self._trials_info[self.get_id(point)] = (point, result)
@property def n_suggested(self): """Number of trials suggested by the algorithm""" return len(self._trials_info) @property def n_observed(self): """Number of completed trials observed by the algorithm""" return sum(bool(point[1] is not None) for point in self._trials_info.values())
[docs] def has_suggested(self, point): """Whether the algorithm has suggested a given point. Parameters ---------- point : tuples of array-likes Points from a `orion.algo.space.Space`. Returns ------- bool True if the point was suggested by the algo, False otherwise. """ return self.get_id(point) in self._trials_info
[docs] def has_observed(self, point): """Whether the algorithm has observed a given point objective. This only counts observed completed trials. Parameters ---------- point : tuples of array-likes Points from a `orion.algo.space.Space`. Returns ------- bool True if the point's objective was observed by the algo, False otherwise. """ trial_id = self.get_id(point) return ( trial_id in self._trials_info and self._trials_info[trial_id][1] is not None )
@property def is_done(self): """Whether the algorithm is done and will not make further suggestions. Return True, if an algorithm holds that there can be no further improvement. By default, the cardinality of the specified search space will be used to check if all possible sets of parameters has been tried. """ if self.n_suggested >= self.space.cardinality: return True if self.n_observed >= getattr(self, "max_trials", float("inf")): return True return False
[docs] def score(self, point): # pylint:disable=no-self-use,unused-argument """Allow algorithm to evaluate `point` based on a prediction about this parameter set's performance. By default, return the same score any parameter (no preference). :returns: A subjective measure of expected perfomance. :rtype: float """ return 0
[docs] def judge(self, point, measurements): # pylint:disable=no-self-use,unused-argument """Inform an algorithm about online `measurements` of a running trial. :param point: A tuple which specifies the values of the (hyper)parameters used to execute user's script with. This method is to be used as a callback in a client-server communication between user's script and a orion's worker using a `BaseAlgorithm`. Data returned from this method must be serializable and will be used as a response to the running environment. Default response is None. .. note:: Calling algorithm to `judge` a `point` based on its online `measurements` will effectively change a state in the algorithm (like a reinforcement learning agent's hidden state or an automatic early stopping mechanism's regression), which it may change the value of the property `should_suspend`. :returns: None or a serializable dictionary containing named data """ return None
@property def should_suspend(self): """Allow algorithm to decide whether a particular running trial is still worth to complete its evaluation, based on information provided by the `judge` method. """ return False @property def configuration(self): """Return tunable elements of this algorithm in a dictionary form appropriate for saving. """ dict_form = dict() for attrname in self._param_names: if attrname.startswith("_"): # Do not log _space or others in conf continue attr = getattr(self, attrname) if isinstance(attr, BaseAlgorithm): attr = attr.configuration dict_form[attrname] = attr return {self.__class__.__name__.lower(): dict_form} @property def space(self): """Domain of problem associated with this algorithm's instance.""" return self._space @space.setter def space(self, space_): """Propagate changes in defined space to possibly nested algorithms.""" self._space = space_ for attr in self.__dict__.values(): if isinstance(attr, BaseAlgorithm): attr.space = space_
# pylint: disable=too-few-public-methods,abstract-method class OptimizationAlgorithm(BaseAlgorithm, metaclass=Factory): """Class used to inject dependency on an algorithm implementation.""" pass