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:
N = self.history.total_num_search
X = self.test.X[self.history.chosen_actions[0:N], :]
t = self.history.fx[0:N]
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)
def write(self, action, t, X=None):
"""
Writing history (update history, not output to a file).
Parameters
----------
action: numpy.ndarray
Indexes of actions.
t: numpy.ndarray
N dimensional array. The negative energy of each search candidate (value of the objective function to be optimized).
X: numpy.ndarray
N x d dimensional matrix. Each row of X denotes the d-dimensional feature vector of each search candidate.
Returns
-------
"""
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 = self.predictor.get_basis(
X) if self.predictor is not None else None
self.history.write(t, action)
self.training.add(X=X, t=t, Z=Z)
if self.new_data is None:
self.new_data = variable(X=X, t=t, Z=Z)
else:
self.new_data.add(X=X, t=t, Z=Z)
def _get_marginal_score(self, mode, chosen_actions, K, alpha):
"""
Getting marginal scores.
Parameters
----------
mode: str
The type of aquision funciton.
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 total number of search candidates.
alpha: float
not used.
Returns
-------
f: list
N dimensional scores (score is defined in each mode)
"""
f = np.zeros((K, len(self.actions)), dtype=float)
new_test_local = self.test.get_subset(chosen_actions)
if self.mpisize == 1:
new_test = new_test_local
else:
new_test = variable()
for nt in self.mpicomm.allgather(new_test_local):
new_test.add(X=nt.X, t=nt.t, Z=nt.Z)
virtual_t = self.predictor.get_predict_samples(self.training, new_test,
K)
for k in range(K):
predictor = copy.deepcopy(self.predictor)
train = copy.deepcopy(self.training)
virtual_train = new_test
virtual_train.t = virtual_t[k, :]
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)
f[k, :] = self.get_score(mode,
predictor=predictor,
training=train,
parallel=False)
return np.mean(f, axis=0)
def __init__(self, test_X, config=None, initial_data=None, comm=None):
"""
Parameters
----------
test_X: numpy.ndarray or physbo.variable
The set of candidates. Each row vector represents the feature vector of each search candidate.
config: set_config object (physbo.misc.set_config)
initial_data: tuple[np.ndarray, np.ndarray]
The initial training datasets.
The first elements is the array of actions and the second is the array of value of objective functions
comm: MPI.Comm, optional
MPI Communicator
"""
self.predictor = None
self.training = variable()
self.new_data = None
self.test = self._make_variable_X(test_X)
self.actions = np.arange(0, self.test.X.shape[0])
self.history = history()
if config is None:
self.config = set_config()
else:
self.config = config
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 len(actions) != len(fs):
msg = "ERROR: len(initial_data[0]) != len(initial_data[1])"
raise RuntimeError(msg)
self.write(actions, fs)
self.actions = 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]
self.config.learning.is_disp = self.mpirank == 0
def _make_variable_X(self, test_X):
"""
Make a new *variable* with X=test_X
Parameters
----------
test_X: numpy.ndarray or physbo.variable
The set of candidates. Each row vector represents the feature vector of each search candidate.
Returns
-------
test_X: numpy.ndarray or physbo.variable
The set of candidates. Each row vector represents the feature vector of each search candidate.
"""
if isinstance(test_X, np.ndarray):
test = variable(X=test_X)
elif isinstance(test_X, variable):
test = test_X
else:
raise TypeError(
"The type of test_X must be ndarray or physbo.variable")
return test