def __init__(self, optimizer, X, Y, args):
'''
Initialize data exporter from initial data (X, Y).
'''
self.optimizer = optimizer
self.problem = optimizer.real_problem
self.n_var, self.n_obj = self.problem.n_var, self.problem.n_obj
self.batch_size = self.optimizer.selection.batch_size
self.iter = 0
self.transformation = optimizer.transformation
# saving path related
self.result_dir = get_result_dir(args)
n_samples = X.shape[0]
# compute initial hypervolume
pfront, pidx = find_pareto_front(Y, return_index=True)
pset = X[pidx]
if args.ref_point is None:
args.ref_point = np.max(Y, axis=0)
hv_value = calc_hypervolume(pfront, ref_point=args.ref_point)
# init data frame
column_names = ['iterID']
d1 = {'iterID': np.zeros(n_samples, dtype=int)}
d2 = {'iterID': np.zeros(len(pset), dtype=int)}
# design variables
for i in range(self.n_var):
var_name = f'x{i + 1}'
d1[var_name] = X[:, i]
d2[var_name] = pset[:, i]
column_names.append(var_name)
# performance
for i in range(self.n_obj):
obj_name = f'f{i + 1}'
d1[obj_name] = Y[:, i]
obj_name = f'Pareto_f{i + 1}'
d2[obj_name] = pfront[:, i]
# predicted performance
for i in range(self.n_obj):
obj_pred_name = f'Expected_f{i + 1}'
d1[obj_pred_name] = np.zeros(n_samples)
obj_pred_name = f'Uncertainty_f{i + 1}'
d1[obj_pred_name] = np.zeros(n_samples)
obj_pred_name = f'Acquisition_f{i + 1}'
d1[obj_pred_name] = np.zeros(n_samples)
d1['Hypervolume_indicator'] = np.full(n_samples, hv_value)
self.export_data = pd.DataFrame(data=d1) # export all data
self.export_pareto = pd.DataFrame(data=d2) # export pareto data
column_names.append('ParetoFamily')
self.export_approx_pareto = pd.DataFrame(columns=column_names) # export pareto approximation data
self.has_family = hasattr(self.optimizer.selection, 'has_family') and self.optimizer.selection.has_family
def save_args(args):
'''
Save arguments to yaml file
'''
result_dir = get_result_dir(args)
args_path = os.path.join(result_dir, 'args.yml')
os.makedirs(os.path.dirname(args_path), exist_ok=True)
with open(args_path, 'w') as f:
yaml.dump(args, f, default_flow_style=False, sort_keys=False)
def experiment_dir(self):
if self.args.resume:
self.expname = os.path.split(self.args.resumedir)[-1]
return self.args.resumedir
elif self.args.eval:
self.expname = os.path.split(self.args.evaldir)[-1]
return self.args.evaldir
expname = utils.get_result_dir(self.base_expdir, self.args.suffix)
self.expname = expname
return os.path.join(self.base_expdir, expname)
def __init__(self, X, Y, args):
'''
Initialize data exporter from initial data (X, Y).
'''
self.n_var, self.n_obj = args.n_var, args.n_obj
self.batch_size = args.batch_size
self.iter = 0
self.X, self.Y = X, Y
self.ref_point = np.max(
Y, axis=0) if args.ref_point is None else args.ref_point
# saving path related
self.result_dir = get_result_dir(args)
n_samples = X.shape[0]
# compute hypervolume
pfront, pidx = find_pareto_front(Y, return_index=True)
pset = X[pidx]
hv_value = calc_hypervolume(pfront, ref_point=self.ref_point)
# init data frame
column_names = ['iterID']
d1 = {'iterID': np.zeros(n_samples, dtype=int)}
d2 = {'iterID': np.zeros(len(pset), dtype=int)}
# design variables
for i in range(self.n_var):
var_name = f'x{i + 1}'
d1[var_name] = X[:, i]
d2[var_name] = pset[:, i]
column_names.append(var_name)
# performance
for i in range(self.n_obj):
obj_name = f'f{i + 1}'
d1[obj_name] = Y[:, i]
obj_name = f'Pareto_f{i + 1}'
d2[obj_name] = pfront[:, i]
d1['Hypervolume_indicator'] = np.full(n_samples, hv_value)
self.export_data = pd.DataFrame(data=d1) # export all data
self.export_pareto = pd.DataFrame(data=d2) # export pareto data
column_names.append('ParetoFamily')
self.export_approx_pareto = pd.DataFrame(
columns=column_names) # export pareto approximation data