# SPDX-License-Identifier: MPL-2.0
# Copyright (C) 2020- The University of Tokyo
#
# This Source Code Form is subject to the terms of the Mozilla Public
# License, v. 2.0. If a copy of the MPL was not distributed with this
# file, You can obtain one at https://mozilla.org/MPL/2.0/.
import numpy as np
import copy
import pickle as pickle
import time
from ._history import History
from ... import _mpi
from .. import discrete
from .. import utility
from .. import score as search_score
from ...misc import SetConfig
from ..._variable import Variable, normalize_t
from ..._rng import make_rng, attach_rng
[ドキュメント]
class Policy(discrete.Policy):
"""Multi objective Bayesian optimization with discrete search space by using unified objective function"""
def __init__(
self,
test_X,
num_objectives,
comm=None,
config=None,
initial_data=None,
rng=None,
):
"""
Initialize the Policy object
Parameters
----------
test_X: numpy.ndarray
The set of candidates. Each row vector represents the feature vector of each search candidate.
num_objectives: int
The number of objectives
comm: MPI.Comm, optional
MPI Communicator
config: physbo.misc.SetConfig, optional
initial_data: tuple[np.ndarray, np.ndarray], optional
The initial training datasets.
The first elements is the array of actions and the second is the array of value of objective functions
"""
self.num_objectives = num_objectives
self.history = History(num_objectives=self.num_objectives)
self.rng = make_rng(rng)
self.training = Variable()
self.training_unified = None
self.predictor = None
self.test = self._make_variable_X(test_X)
self.new_data = None
self.actions = np.arange(0, test_X.shape[0])
if config is None:
self.config = SetConfig()
else:
self.config = config
self.ard = False
if initial_data is not None:
if len(initial_data) != 2:
msg = "ERROR: initial_data should be 2-elements tuple or list (actions and objectives)"
raise RuntimeError(msg)
actions, fs = initial_data
if fs.shape[1] != self.num_objectives:
msg = "ERROR: initial_data[1].shape[1] != num_objectives"
raise RuntimeError(msg)
if len(actions) != fs.shape[0]:
msg = "ERROR: len(initial_data[0]) != initial_data[1].shape[0]"
raise RuntimeError(msg)
self.write(actions, fs)
self.actions = np.array(sorted(list(set(self.actions) - set(actions))))
if comm is None:
self.mpicomm = None
self.mpisize = 1
self.mpirank = 0
else:
self.mpicomm = comm
self.mpisize = comm.size
self.mpirank = comm.rank
self.actions = np.array_split(self.actions, self.mpisize)[self.mpirank]
[ドキュメント]
def write(
self,
action,
t,
X=None,
time_total=None,
time_update_predictor=None,
time_get_action=None,
time_run_simulator=None,
):
self.history.write(
t,
action,
time_total=time_total,
time_update_predictor=time_update_predictor,
time_get_action=time_get_action,
time_run_simulator=time_run_simulator,
)
action = np.array(action)
N = len(action)
t = np.array(t)
# Ensure t is 2D: shape (N, num_objectives)
if t.ndim == 1:
if N == 1:
t = t.reshape(1, -1)
else:
raise ValueError(f"Number of actions is {N} > 1, but t is 1D array")
assert action.shape[0] == t.shape[0], "The number of actions and t must be the same"
assert t.shape[1] == self.num_objectives, "The number of objectives in t must be the same as num_objectives"
# Determine X and Z (different for each objective)
if X is None:
X = self.test.X[action, :]
Z = self.test.Z[:, action, :] if self.test.Z is not None else None
else:
Z = None
# Failed observations (any non-finite objective) are kept in the
# history and their actions are consumed, but they are not used
# for training.
valid = utility.finite_mask(t)
if np.any(valid):
X_ok = np.asarray(X)[valid]
t_ok = t[valid]
Z_ok = utility.mask_rows(Z, valid)
if self.new_data is None:
self.new_data = Variable(X=X_ok, t=t_ok, Z=Z_ok)
else:
self.new_data.add(X=X_ok, t=t_ok, Z=Z_ok)
# Add to single training Variable with full 2D t matrix and (k, N, n) Z
if self.training.X is None:
self.training = Variable(X=X_ok, t=t_ok, Z=Z_ok)
else:
self.training.add(X=X_ok, t=t_ok, Z=Z_ok)
# remove action from candidates if exists
if len(self.actions) > 0:
local_index = np.searchsorted(self.actions, action)
local_index = local_index[
np.take(self.actions, local_index, mode="clip") == action
]
self.actions = self._delete_actions(local_index)
[ドキュメント]
def random_search(
self,
max_num_probes,
num_search_each_probe=1,
simulator=None,
is_disp=True,
disp_pareto_set=False,
):
if self.mpirank != 0:
is_disp = False
N = int(num_search_each_probe)
if is_disp:
utility.show_interactive_mode(simulator, self.history)
for n in range(0, max_num_probes):
time_total = time.time()
if is_disp and N > 1:
utility.show_start_message_multi_search(
self.history.num_runs, score="random"
)
time_get_action = time.time()
action = self._get_random_action(N)
time_get_action = time.time() - time_get_action
if simulator is None:
return action
time_run_simulator = time.time()
t = _run_simulator(simulator, action, self.mpicomm)
time_run_simulator = time.time() - time_run_simulator
time_total = time.time() - time_total
self.write(
action,
t,
time_total=[time_total] * N,
time_update_predictor=np.zeros(N, dtype=float),
time_get_action=[time_get_action] * N,
time_run_simulator=[time_run_simulator] * N,
)
if is_disp:
utility.show_search_results_mo(
self.history, N, disp_pareto_set=disp_pareto_set
)
return copy.deepcopy(self.history)
[ドキュメント]
def bayes_search(
self,
training_list=None,
max_num_probes=None,
num_search_each_probe=1,
predictor=None,
is_disp=True,
disp_pareto_set=False,
simulator=None,
score="EI",
interval=0,
num_rand_basis=0,
unify_method=None,
optimizer=None,
ard=False,
ucb_beta=1.0,
):
"""
Performing Bayesian optimization by using unified objective function
Parameters
----------
training_list: list of physbo.Variable, optional
The training datasets.
max_num_probes: int, optional
The maximum number of searching process by Bayesian optimization.
num_search_each_probe: int, optional
The number of searching by Bayesian optimization at each process.
predictor: predictor object, optional
The predictor object.
is_disp: bool, optional
If true, process messages are outputted.
disp_pareto_set: bool, optional
If true, Pareto set is displayed.
simulator: callable, optional
The simulator function.
score: str or callable
TS, EI, PI, UCB, or a callable called positionally as
``score(predictor, training, test)`` returning a numpy.ndarray of
shape (test.X.shape[0],), with larger values preferred. Empty test
data skips the callable. Under MPI, the callable evaluates local
points and must not perform collective communication.
training contains unified targets of shape (N, 1).
interval: int, optional
The interval number of learning the hyper parameter.
If you set the negative value to interval, the hyper parameter learning is not performed.
If you set zero to interval, the hyper parameter learning is performed only at the first step.
num_rand_basis: int, optional
The number of basis function. If you choose 0, ordinary Gaussian process run.
unify_method: callable
The unified objective function. It is a function or a callable object that maps a (N, num_objectives) numpy.ndarray of original objective functions to a (N, 1) numpy.ndarray of unified objective functions.
See physbo.search.unify for examples.
If it draws random numbers through an ``rng`` attribute (as
``ParEGO`` does) and was not given a generator of its own, it
uses the random number generator of the policy.
optimizer: Optimizer object, optional
This is for compatibility with the range-based Policies.
This is not used.
ard: bool, optional
If True, use Automatic Relevance Determination (ARD) for the Gaussian kernel.
Default is False.
ucb_beta: float
Trade-off parameter between exploitation and exploration used for
string "UCB". The score is ``fmean + ucb_beta * fstd``.
Default is 1.0; ignored for callable scores.
Returns
-------
history: history object (physbo.search.discrete_unified.results.history)
"""
score = search_score.resolve(
score, alpha=self.config.search.alpha, ucb_beta=ucb_beta, rng=self.rng
)
assert unify_method is not None, "unify_method must be provided"
if self.mpirank != 0:
is_disp = False
old_disp = self.config.learning.is_disp
self.config.learning.is_disp = is_disp
if max_num_probes is None:
max_num_probes = 1
simulator = None
self.unify_method = unify_method
attach_rng(unify_method, self.rng)
if num_rand_basis < 0:
raise ValueError("num_rand_basis must be non-negative")
is_rand_expans = (num_rand_basis > 0)
self.ard = ard
if training_list is not None:
self.training = training_list
if predictor is None:
self._initialize_predictor(is_rand_expans)
else:
self.predictor = predictor
if max_num_probes > 0 or interval >= 0:
utility.require_training_data(self.training)
if max_num_probes == 0 and interval >= 0:
self._learn_hyperparameter(num_rand_basis)
N = int(num_search_each_probe)
for n in range(max_num_probes):
time_total = time.time()
time_update_predictor = time.time()
if utility.is_learning(n, interval):
self._learn_hyperparameter(num_rand_basis)
else:
self._update_predictor()
time_update_predictor = time.time() - time_update_predictor
if num_search_each_probe != 1:
utility.show_start_message_multi_search(
self.history.num_runs, utility.score_name(score)
)
time_get_action = time.time()
K = self.config.search.multi_probe_num_sampling
alpha = self.config.search.alpha
action = self._get_actions(score, N, K, alpha, ucb_beta)
time_get_action = time.time() - time_get_action
N_indeed = len(action)
if N_indeed == 0:
if self.mpirank == 0:
print("WARNING: All actions have already searched.")
self.config.learning.is_disp = old_disp
return copy.deepcopy(self.history)
if simulator is None:
self.config.learning.is_disp = old_disp
return action
time_run_simulator = time.time()
t = _run_simulator(simulator, action, self.mpicomm)
time_run_simulator = time.time() - time_run_simulator
time_total = time.time() - time_total
self.write(
action,
t,
time_total=[time_total] * N_indeed,
time_update_predictor=[time_update_predictor] * N_indeed,
time_get_action=[time_get_action] * N_indeed,
time_run_simulator=[time_run_simulator] * N_indeed,
)
if is_disp:
utility.show_search_results_mo(
self.history, N, disp_pareto_set=disp_pareto_set
)
self._update_predictor()
self.config.learning.is_disp = old_disp
return copy.deepcopy(self.history)
def _get_actions(self, mode, N, K, alpha, ucb_beta=1.0):
self._update_predictor()
comm = self.mpicomm if self.mpisize > 1 else None
# TS: one posterior sample for every process, drawn on rank 0
first_mode = search_score.share_sample(
mode, self.predictor, self.training_unified, comm
)
f, error = _mpi.call(
self.mpicomm,
self.get_score,
mode=first_mode,
alpha=alpha,
ucb_beta=ucb_beta,
parallel=False,
)
champion, local_champion, local_index = self._find_champion(f, error)
if champion == -1:
return np.zeros(0, dtype=int)
if champion == local_champion:
self.actions = self._delete_actions(local_index)
chosen_actions = [champion]
for n in range(1, N):
f, error = self._get_marginal_score(
mode, chosen_actions[0:n], K, alpha, ucb_beta
)
champion, local_champion, local_index = self._find_champion(f, error)
if champion == -1:
break
if champion == local_champion:
self.actions = self._delete_actions(local_index)
chosen_actions.append(champion)
return np.array(chosen_actions)
[ドキュメント]
def get_post_fmean(self, xs):
"""
Calculate mean value of predictor (post distribution)
Parameters
----------
xs: physbo.Variable or np.ndarray
input parameters to calculate mean value
shape is (num_points, num_parameters)
Returns
-------
fmean: numpy.ndarray
Mean value of the post distribution.
Returned shape is (num_points).
"""
X = self._make_variable_X(xs)
if self.predictor is None:
self._warn_no_predictor("get_post_fmean()")
predictor = self._make_gp_predictor()
predictor.fit(self.training_unified, 0, comm=self.mpicomm, rng=self.rng)
predictor.prepare(self.training_unified)
return predictor.get_post_fmean(self.training_unified, X)
else:
self._update_predictor()
return self.predictor.get_post_fmean(self.training_unified, X)
[ドキュメント]
def get_post_fcov(self, xs, diag=True):
"""
Calculate covariance of predictor (post distribution)
Parameters
----------
xs: physbo.Variable or np.ndarray
input parameters to calculate covariance
shape is (num_points, num_parameters)
diag: bool
If true, only variances (diagonal elements) are returned.
Returns
-------
fcov: numpy.ndarray
Covariance matrix of the post distribution.
Returned shape is (num_points) if diag=true, (num_points, num_points) if diag=false.
"""
X = self._make_variable_X(xs)
if self.predictor is None:
self._warn_no_predictor("get_post_fcov()")
predictor = self._make_gp_predictor()
predictor.fit(self.training_unified, 0, comm=self.mpicomm, rng=self.rng)
predictor.prepare(self.training_unified)
return predictor.get_post_fcov(self.training_unified, X, diag)
else:
self._update_predictor()
return self.predictor.get_post_fcov(self.training_unified, X, diag)
[ドキュメント]
def get_score(
self,
mode,
*,
actions=None,
xs=None,
predictor=None,
training=None,
parallel=True,
alpha=1,
ucb_beta=1.0,
):
"""
Calculate score (acquisition function)
Parameters
----------
mode: str or callable
TS, EI, PI, UCB, or a callable called positionally as
``mode(predictor, training, test)`` returning a numpy.ndarray of
shape (test.X.shape[0],), with larger values preferred. Empty test
data skips the callable but still participates in MPI gathering.
The callable must not perform collective communication. alpha and
ucb_beta apply only to string modes.
training contains unified targets of shape (N, 1).
actions: array of int
actions to calculate score
xs: physbo.Variable or np.ndarray
input parameters to calculate score
predictor: predictor object
predictor used to calculate score.
If not given, self.predictor will be used.
training:physbo.Variable
Training dataset.
If not given, self.training will be used.
parallel: bool
Calculate scores in parallel by MPI (default: True)
alpha: float
Tuning parameter which is used if mode = TS.
In TS, multi variation is tuned as np.random.multivariate_normal(mean, cov*alpha**2, size).
ucb_beta: float
Trade-off parameter between exploitation and exploration used for
string "UCB". The score is ``fmean + ucb_beta * fstd``.
Default is 1.0; ignored for callable scores.
Returns
-------
f: float or list of float
Score defined in each mode.
Raises
------
RuntimeError
If both *actions* and *xs* are given
Notes
-----
When neither *actions* nor *xs* are given, scores for actions not yet searched will be calculated.
When *parallel* is True, it is assumed that the function receives the same input (*actions* or *xs*) for all the ranks.
If you want to split the input array itself, set *parallel* be False and merge results by yourself.
"""
mode = search_score.resolve(
mode, alpha=alpha, ucb_beta=ucb_beta, rng=self.rng
)
if training is None:
training = self.training_unified
if training.X is None or training.X.shape[0] == 0:
msg = "ERROR: No training data is registered."
raise RuntimeError(msg)
if predictor is None:
if self.predictor is None:
self._warn_no_predictor("get_score()")
predictor = self._make_gp_predictor()
predictor.fit(training, 0, comm=self.mpicomm, rng=self.rng)
predictor.prepare(training)
else:
self._update_predictor()
predictor = self.predictor
def scores(actions):
# the share of this process of the candidates
if xs is not None:
if actions is not None:
raise RuntimeError("ERROR: both actions and xs are given")
test = self._make_variable_X(xs)
if parallel and self.mpisize > 1:
actions = np.array_split(
np.arange(test.X.shape[0]), self.mpisize
)
test = test.get_subset(actions[self.mpirank])
else:
if actions is None:
actions = self.actions
else:
if isinstance(actions, int):
actions = [actions]
if parallel and self.mpisize > 1:
actions = np.array_split(actions, self.mpisize)[
self.mpirank
]
test = self.test.get_subset(actions)
return search_score.score(
mode,
predictor=predictor,
training=training,
test=test,
alpha=alpha,
ucb_beta=ucb_beta,
)
comm = self.mpicomm if parallel and self.mpisize > 1 else None
# TS: one posterior sample for every process, drawn on rank 0
mode = search_score.share_sample(mode, predictor, training, comm)
f, error = _mpi.call(comm, scores, actions)
if comm is not None:
fs = _mpi.allgather(comm, f, error)
f = np.hstack(fs)
return f
[ドキュメント]
def get_permutation_importance(self, n_perm: int, split_features_parallel=False):
"""
Calculating permutation importance of model
Parameters
==========
n_perm: int
number of permutations
split_features_parallel: bool
If true, split features in parallel.
Returns
=======
importance_mean: numpy.ndarray
importance_mean
importance_std: numpy.ndarray
importance_std
"""
if self.predictor is None:
self._warn_no_predictor("get_permutation_importance()")
predictor = self._make_gp_predictor()
predictor.fit(self.training_unified, 0, rng=self.rng)
predictor.prepare(self.training_unified)
return predictor.get_permutation_importance(
self.training_unified,
n_perm,
comm=self.mpicomm,
split_features_parallel=split_features_parallel,
rng=self.rng,
)
else:
self._update_predictor()
return self.predictor.get_permutation_importance(
self.training_unified,
n_perm,
comm=self.mpicomm,
split_features_parallel=split_features_parallel,
rng=self.rng,
)
def _get_marginal_score(self, mode, chosen_actions, K, alpha, ucb_beta=1.0):
"""
Getting marginal scores.
Parameters
----------
mode: str
The type of acquisition function.
TS (Thompson Sampling), EI (Expected Improvement) and PI (Probability of Improvement) are available.
These functions are defined in score.py.
chosen_actions: numpy.ndarray
Array of selected actions.
K: int
The number of samples for evaluating score.
alpha: float
not used.
Returns
-------
f: numpy.ndarray or None
The marginal scores of the candidates of this process (score is
defined in each mode), or None if ``error`` is not None.
error: Exception or None
An exception raised on this process while computing ``f`` in a
parallel run, to be passed to :meth:`_find_champion` (in a serial
run it is raised).
"""
comm = self.mpicomm if self.mpisize > 1 else None
# Draw K samples of the values of objective function of chosen actions.
# self.test is not partitioned over ranks, so the chosen points are
# identical on every rank; the virtual values are drawn on rank 0 and
# broadcast so that all ranks fantasize the same observations.
new_test = self.test.get_subset(chosen_actions)
if comm is None or self.mpirank == 0:
virtual_t, error = _mpi.call(
comm,
self.predictor.get_predict_samples,
self.training_unified,
new_test,
K,
rng=self.rng,
)
else:
virtual_t, error = None, None
if comm is not None:
virtual_t = _mpi.bcast(comm, virtual_t, error, root=0)
def model(k):
# the predictor and the training data with the k-th virtual
# observations added (the same on every process)
predictor = copy.deepcopy(self.predictor)
train = copy.deepcopy(self.training_unified)
virtual_train = new_test
# Normalize virtual_t[k, :] to (N, 1) shape
virtual_train.t = normalize_t(virtual_t[k, :], k=1)
if virtual_train.Z is None:
train.add(virtual_train.X, virtual_train.t)
else:
train.add(virtual_train.X, virtual_train.t, virtual_train.Z)
predictor.update(train, virtual_train)
return predictor, train
# TS: one posterior sample per virtual data set for every process,
# drawn on rank 0
samples = search_score.share_samples(mode, self.predictor, model, K, comm)
return _mpi.call(comm, self._marginal_score, mode, model, K, ucb_beta, samples)
def _marginal_score(self, mode, model, K, ucb_beta, samples=None):
"""The marginal scores of this process's candidates: the mean of the
scores over the K models ``model(k)`` of :meth:`_get_marginal_score`.
``samples`` are their posterior samples for Thompson sampling (see
:func:`physbo.search.score.share_samples`), or None."""
f = np.zeros((K, len(self.actions)), dtype=float)
for k in range(K):
predictor, train = model(k)
mode_k = mode if samples is None else mode.with_sample(samples[k])
f[k, :] = self.get_score(
mode_k,
predictor=predictor,
training=train,
ucb_beta=ucb_beta,
parallel=False,
)
return np.mean(f, axis=0)
[ドキュメント]
def save(self, file_history, file_training=None, file_predictor=None):
"""
Saving history, training and predictor into the corresponding files.
Parameters
----------
file_history: str
The name of the file that stores the information of the history.
file_training: str
The name of the file that stores the training dataset.
file_predictor: str
The name of the file that stores the predictor dataset.
Returns
-------
"""
if self.mpirank == 0:
self.history.save(file_history)
if file_training is not None:
self.training.save(file_training)
if file_predictor is not None:
with open(file_predictor, "wb") as f:
pickle.dump(self.predictor, f)
[ドキュメント]
def load(self, file_history, file_training=None, file_predictor=None):
"""
Loading files about history, training and predictor.
Parameters
----------
file_history: str
The name of the file that stores the information of the history.
file_training: str
The name of the file that stores the training dataset.
file_predictor: str
The name of the file that stores the predictor dataset.
Returns
-------
"""
self.history.load(file_history)
if file_training is None:
# rebuild the training data from the valid observations only
actions, t = self.history.export_valid()
X = self.test.X[actions, :]
self.training = Variable(X=X, t=t)
else:
self.training = Variable()
self.training.load(file_training)
if file_predictor is not None:
with open(file_predictor, "rb") as f:
self.predictor = pickle.load(f)
N = self.history.total_num_search
visited = self.history.chosen_actions[:N]
local_index = np.searchsorted(self.actions, visited)
local_index = local_index[
np.take(self.actions, local_index, mode="clip") == visited
]
self.actions = self._delete_actions(local_index)
def _learn_hyperparameter(self, num_rand_basis):
self.training_unified = self._unify_training(self.training)
self.predictor.fit(
self.training_unified, num_rand_basis, comm=self.mpicomm, rng=self.rng
)
self.predictor.prepare(self.training_unified)
Z = self.predictor.get_basis(self.training_unified.X)
if Z is not None:
self.training_unified.Z = Z[np.newaxis, :, :]
self.new_data = None
def _initialize_predictor(self, is_rand_expans):
if is_rand_expans:
self.predictor = self._make_blm_predictor()
else:
self.predictor = self._make_gp_predictor()
def _update_predictor(self):
if self.new_data is not None:
self.training_unified = self._unify_training(self.training)
N = self.training_unified.t.shape[0]
n = self.new_data.t.shape[0]
new_data_unified = self.training_unified.get_subset(np.arange(N - n, N))
assert np.allclose(new_data_unified.X, self.new_data.X)
self.predictor.update(self.training_unified, new_data_unified)
self.new_data = None
def _unify_training(self, training: Variable) -> Variable:
"""
Wrapper of the unify_method function
"""
t_unified = self.unify_method(training.t)
return Variable(X=training.X, t=t_unified.reshape(-1, 1))
def _run_simulator(simulator, action, comm=None):
if comm is None:
return simulator(action)
if comm.rank == 0:
t, error = _mpi.call(comm, simulator, action)
else:
t, error = 0.0, None
return _mpi.bcast(comm, t, error, root=0)