def newCandidate_beyes(self):
N_newCandidates = self.Ncand_min
# Number of new candidated made by each core
Neach = int(np.ceil(N_newCandidates / self.comm.size))
# Use all cores on nodes.
N_newCandidates = Neach * N_newCandidates
# perform mutations
self.comm.barrier()
if self.master:
print('mutations starting', flush=True)
t0 = time()
anew_mutated_list = self.mutate(Neach)
self.comm.barrier()
if self.master:
print('mutation time:', time() - t0, flush=True)
# Relax with MLmodel
anew_list = []
E_list = []
error_list = []
for anew_mutated in anew_mutated_list:
anew = self.relaxML(anew_mutated, with_error=self.errorRelax)
anew_list.append(anew)
E, error, theta0 = self.MLmodel.predict_energy(anew,
return_error=True)
E_list.append(E)
error_list.append(error)
E_list = np.array(E_list)
error_list = np.array(error_list)
self.comm.barrier()
if self.master:
print('sqrt(theta0):', np.sqrt(np.abs(theta0)), flush=True)
operation_index = np.array([
self.operation_dict.index(a.info['key_value_pairs']['origin'])
for a in anew_list
]).astype(int)
# Syncronize all new candidates on all cores
anew_mutated_all = sync_atoms(world,
atoms_list=anew_mutated_list,
operation_dict=self.operation_dict,
operation_index=operation_index)
anew_all = sync_atoms(world,
atoms_list=anew_list,
Epred_list=E_list,
error_list=error_list,
kappa=self.kappa,
operation_dict=self.operation_dict,
operation_index=operation_index)
self.comm.barrier()
if self.master:
print('candidates syncronized', flush=True)
# filter away very uncertain structures
error_all_tmp = np.array(
[a.info['key_value_pairs']['predictedError'] for a in anew_all])
min_certainty = self.min_certainty
for _ in range(5):
filt = error_all_tmp < min_certainty * np.sqrt(np.abs(theta0))
if np.sum(filt.astype(int)) > 0:
anew_mutated_all = [
anew_mutated_all[i] for i in range(len(anew_all))
if filt[i]
]
anew_all = [
anew_all[i] for i in range(len(anew_all)) if filt[i]
]
break
else:
min_certainty = min_certainty + (1 - min_certainty) / 2
# include relaxed population
if (self.NsearchIter % self.Nuse_pop_as_candidates) == 0:
anew_mutated_all = self.population.pop + anew_mutated_all
anew_all = self.population.pop_MLrelaxed + anew_all
self.comm.barrier()
if self.master:
print('model relaxed candidates done', flush=True)
# Write candidates to file
if self.master:
label_relaxed = self.ML_dir + 'ML_relaxed{}'.format(
self.traj_counter)
write(label_relaxed + '.traj', anew_all, parallel=False)
label_unrelaxed = self.ML_dir + 'ML_unrelaxed{}'.format(
self.traj_counter)
write(label_unrelaxed + '.traj', anew_mutated_all, parallel=False)
# Add force-mutated structures to candidates
#anew_forceMut, anew_preForceMut = self.get_force_mutated_population()
#anew_all += anew_forceMut
#anew_mutated_all += anew_preForceMut
# Extract + print data
E_all = np.array(
[a.info['key_value_pairs']['predictedEnergy'] for a in anew_all])
error_all = np.array(
[a.info['key_value_pairs']['predictedError'] for a in anew_all])
fitness_all = np.array(
[a.info['key_value_pairs']['fitness'] for a in anew_all])
if self.master:
print('{}:\n'.format(self.traj_counter), np.c_[E_all, error_all,
fitness_all])
return anew_all, anew_mutated_all, E_all, error_all
def newCandidate_beyes(self):
N_newCandidates = self.Ncand_min
# Number of new candidated made by each core
Neach = int(np.ceil(N_newCandidates / self.comm.size))
# Use all cores on nodes.
N_newCandidates = Neach * N_newCandidates
# perform mutations
if self.master:
t0 = time()
anew_mutated_list = self.mutate(Neach)
if self.master:
print('mutation time:', time() - t0, flush=True)
# Relax with MLmodel
anew_list = []
E_list = []
error_list = []
for anew_mutated in anew_mutated_list:
anew = self.relaxML(anew_mutated, with_error=True)
anew_list.append(anew)
E, error, theta0 = self.MLmodel.predict_energy(anew,
return_error=True)
E_list.append(E)
error_list.append(error)
E_list = np.array(E_list)
error_list = np.array(error_list)
if self.master:
print('theta0:', theta0, flush=True)
operation_index = np.array([
self.operation_dict.index(a.info['key_value_pairs']['origin'])
for a in anew_list
]).astype(int)
# Syncronize all new candidates on all cores
anew_mutated_all = sync_atoms(world,
atoms_list=anew_mutated_list,
operation_dict=self.operation_dict,
operation_index=operation_index)
anew_all = sync_atoms(world,
atoms_list=anew_list,
Epred_list=E_list,
error_list=error_list,
kappa=self.kappa,
operation_dict=self.operation_dict,
operation_index=operation_index)
error_all_tmp = np.array(
[a.info['key_value_pairs']['predictedError'] for a in anew_all])
min_certainty = self.min_certainty
for _ in range(5):
filt = error_all_tmp < min_certainty * np.sqrt(np.abs(theta0))
if np.sum(filt.astype(int)) > 0:
anew_mutated_all = [
anew_mutated_all[i] for i in range(len(anew_all))
if filt[i]
]
anew_all = [
anew_all[i] for i in range(len(anew_all)) if filt[i]
]
break
else:
min_certainty = min_certainty + (1 - min_certainty) / 2
# Write candidates to file
if self.master:
label_relaxed = self.ML_dir + 'ML_relaxed{}'.format(
self.traj_counter)
write(label_relaxed + '.traj', anew_all, parallel=False)
label_unrelaxed = self.ML_dir + 'ML_unrelaxed{}'.format(
self.traj_counter)
write(label_unrelaxed + '.traj', anew_mutated_all, parallel=False)
# Extract + print data
E_all = np.array(
[a.info['key_value_pairs']['predictedEnergy'] for a in anew_all])
error_all = np.array(
[a.info['key_value_pairs']['predictedError'] for a in anew_all])
fitness_all = np.array(
[a.info['key_value_pairs']['fitness'] for a in anew_all])
if self.master:
print('{}:\n'.format(self.traj_counter), np.c_[E_all, error_all,
fitness_all])
return anew_all, anew_mutated_all, E_all, error_all
def newCandidate_beyes(self, prior_fitness=False):
N_newCandidates = self.Ncand_min
# Number of new candidated made by each core
Neach = int(np.ceil(N_newCandidates / self.comm.size))
# Use all cores on nodes.
N_newCandidates = Neach * N_newCandidates
# perform mutations
if self.master:
t0 = time()
anew_mutated_list = self.mutate(Neach)
if self.master:
print('mutation time:', time() - t0, flush=True)
# Relax with MLmodel
anew_list = []
E_list = []
error_list = []
dmin_list = []
for anew_mutated in anew_mutated_list:
anew = self.relaxML(anew_mutated, with_error=self.errorRelax)
anew_list.append(anew)
if prior_fitness:
E, error, theta0 = self.MLmodel_prior.predict_energy(
anew, return_error=True)
else:
E, error, theta0 = self.MLmodel.predict_energy(
anew, return_error=True)
E_list.append(E)
error_list.append(error)
dmin = self.get_minDistance2data(anew)
dmin_list.append(dmin)
E_list = np.array(E_list)
error_list = np.array(error_list)
dmin_list = np.array(dmin_list)
if self.master:
if self.use_fine_model:
print('sqrt(theta0_prior):',
np.sqrt(np.abs(self.MLmodel_prior.theta0)),
'sqrt(theta0_fine):',
np.sqrt(np.abs(self.MLmodel_fine.theta0)),
flush=True)
else:
print('sqrt(theta0_prior):',
np.sqrt(np.abs(self.MLmodel_prior.theta0)),
flush=True)
operation_index = np.array([
self.operation_dict.index(a.info['key_value_pairs']['origin'])
for a in anew_list
]).astype(int)
# Syncronize all new candidates on all cores
anew_mutated_all = sync_atoms(world,
atoms_list=anew_mutated_list,
operation_dict=self.operation_dict,
operation_index=operation_index)
anew_all = sync_atoms(world,
atoms_list=anew_list,
Epred_list=E_list,
error_list=error_list,
dmin_list=dmin_list,
kappa=self.kappa,
operation_dict=self.operation_dict,
operation_index=operation_index)
# Filter out very uncertain structures
theta0_prior = self.MLmodel_prior.theta0
error_all_tmp = np.array(
[a.info['key_value_pairs']['predictedError'] for a in anew_all])
min_certainty = self.min_certainty
for _ in range(5):
filt = error_all_tmp < min_certainty * np.sqrt(
np.abs(theta0_prior))
if np.sum(filt.astype(int)) > 0:
anew_mutated_all = [
anew_mutated_all[i] for i in range(len(anew_all))
if filt[i]
]
anew_all = [
anew_all[i] for i in range(len(anew_all)) if filt[i]
]
break
else:
min_certainty = min_certainty + (1 - min_certainty) / 2
# Filter structures that are too close to known data
"""
if self.use_fine_model:
min_distance = 0.02 * self.MLmodel_fine.sigma
else:
min_distance = self.min_distance
"""
dmin_all_tmp = np.array(
[a.info['key_value_pairs']['dmin'] for a in anew_all])
distance_filter = dmin_all_tmp > self.min_distance
anew_mutated_all = [
anew_mutated_all[i] for i in range(len(anew_all))
if distance_filter[i]
]
anew_all = [
anew_all[i] for i in range(len(anew_all)) if distance_filter[i]
]
self.comm.barrier()
if self.master:
print('model relaxed candidates done', flush=True)
# Write candidates to file
if self.master:
label_relaxed = self.ML_dir + 'ML_relaxed{}'.format(
self.traj_counter)
write(label_relaxed + '.traj', anew_all, parallel=False)
label_unrelaxed = self.ML_dir + 'ML_unrelaxed{}'.format(
self.traj_counter)
write(label_unrelaxed + '.traj', anew_mutated_all, parallel=False)
# Add force-mutated structures to candidates
#anew_forceMut, anew_preForceMut = self.get_force_mutated_population()
#anew_all += anew_forceMut
#anew_mutated_all += anew_preForceMut
# Extract + print data
E_all = np.array(
[a.info['key_value_pairs']['predictedEnergy'] for a in anew_all])
error_all = np.array(
[a.info['key_value_pairs']['predictedError'] for a in anew_all])
fitness_all = np.array(
[a.info['key_value_pairs']['fitness'] for a in anew_all])
if self.master:
print('{}:\n'.format(self.traj_counter), np.c_[E_all, error_all,
fitness_all])
return anew_all, anew_mutated_all, E_all, error_all