Skip to content

Commit

Permalink
update population initialization
Browse files Browse the repository at this point in the history
  • Loading branch information
@172698691
172698691 authored Nov 1, 2022
1 parent 703466a commit 1bf9968
Showing 1 changed file with 33 additions and 129 deletions.
162 changes: 33 additions & 129 deletions main.py
View file
Original file line number Diff line number Diff line change
@@ -1,7 +1,6 @@
# Unified MAML for this project
from __future__ import annotations
from cProfile import label
from copy import deepcopy

from typing import Tuple, List

Expand All @@ -21,7 +20,6 @@
from pymoo.indicators.igd import IGD

from utils import NamedDict
from examples import train_baseline_nn
from examples.example import get_args, get_network_structure, get_dataset
from examples.example_sinewave import get_args_maml_regression, get_network_structure_maml_regression, \
get_dataset_sinewave
Expand Down Expand Up @@ -180,28 +178,21 @@ def __call__(self, x: np.ndarray) -> List[float]:


class MyProblem(Problem):
def __init__(self, sol: Sol, min_max: Tuple[int, int] | None):
def __init__(self, sol: Sol):
self.sol = sol
args = sol.args
n_var, n_obj = args.problem_dim
if min_max is None:
min_max = (0, 1)
self.min_val, self.max_val = min_max
super().__init__(n_var=n_var, # 变量数
n_obj=n_obj, # 目标数
# n_constr=2, # 约束数
xl=np.array([0] * n_var, np.float32), # 变量下界
xu=np.array([1] * n_var, np.float32), # 变量上界
super().__init__(n_var=8,
n_obj=3,
# n_constr=2,
xl=np.array([0] * 8, np.float32),
xu=np.array([1] * 8, np.float32),
)

def _evaluate(self, x, out, *args, **kwargs):
x = x.astype(np.float32)
# 定义目标函数
f = []
for xi in x:
fi = self.sol(xi)
f.append(fi)
# todo
out["F"] = np.array(f)


Expand Down Expand Up @@ -254,13 +245,11 @@ def main_NSGA():
network_structure = get_network_structure(args)
# generate delta
delta = []
# for i in range(2):
# delta.append([np.random.randint(0, 100, args.train_test[i]), np.random.randint(0, 10, args.train_test[i])])
delta.append([np.random.randint(0, 100, args.train_test[0]), np.random.randint(0, 10, args.train_test[0])])
delta.append([np.array([np.mean(delta[0][0])]), np.array([np.mean(delta[0][1])])])
for i in range(2):
delta.append([np.random.randint(0, 100, args.train_test[i]), np.random.randint(0, 10, args.train_test[i])])
dataset, min_max = get_dataset(args, normalize_targets=True, delta=delta)
sol = Sol(dataset, args, network_structure)
train_loss = sol.train(explicit=False)
# train_loss = sol.train(explicit=False)
test_loss = sol.test(return_single_loss=False)

n_var = args.problem_dim[0]
Expand All @@ -269,22 +258,27 @@ def main_NSGA():

pf_true = get_pf(n_var, n_objectives, delta_finetune, min_max)

history_x, history_f = np.empty((0, n_var)), np.empty((0, n_objectives))
igd = []
x_size = []
fn_eval_limit = 600
pop_size = 50
n_gen = int(fn_eval_limit / pop_size)
max_pts_num = 10

saved_x = None
saved_y = None

fn_eval_limit = 300
max_pts_num = 5
pop_size = 60
sample_size = 30
n_gen = 10

while sum(x_size) < fn_eval_limit:

algorithm = NSGA2(pop_size=pop_size)
if history_x.shape[0] < sample_size:
random_x = np.random.rand(pop_size - history_x.shape[0], n_var)
sample_x = np.vstack((history_x, random_x))
else:
random_x = np.random.rand(pop_size - sample_size, n_var)
sample_x = np.vstack((history_x[np.random.choice(history_x.shape[0], sample_size)], random_x))

algorithm = NSGA2(pop_size=pop_size, sampling=sample_x)

res = minimize(MyProblem(sol=sol, min_max=min_max),
res = minimize(MyProblem(sol=sol),
algorithm,
("n_gen", n_gen),
seed=1,
Expand All @@ -293,14 +287,13 @@ def main_NSGA():
X = res.X
if len(X) > max_pts_num:
X = X[np.random.choice(X.shape[0], max_pts_num)]

history_x = np.vstack((history_x, X))
history_f = np.vstack((history_f, res.F))

x_size.append(X.shape[0])

X = X.astype(np.float32)
if saved_x is None:
saved_x = X
else:
saved_x = np.concatenate((saved_x, X))
y_true = evaluate(X, delta_finetune, n_objectives, min_max=min_max)

new_y_true = []
Expand All @@ -309,19 +302,13 @@ def main_NSGA():
new_y_true = np.array(new_y_true, dtype=np.float32)
new_y_true = new_y_true.reshape((*new_y_true.shape, 1))

if saved_y is None:
saved_y = new_y_true
else:
saved_y = np.concatenate((saved_y, new_y_true), axis=1)

# sol.test_continue(X, new_y_true)
sol.test_continue(saved_x, saved_y)
sol.test_continue(X, new_y_true)

metric = IGD(pf_true, zero_to_one=True)
igd.append(metric.do(res.F))
igd.append(metric.do(history_f))

pf = evaluate(res.X, delta_finetune, n_objectives, min_max=min_max)
moea_pf, n_evals_moea, igd_moea = get_moea_data(n_var, n_objectives, delta_finetune, algorithm, n_gen, min_max)
moea_pf, n_evals_moea, igd_moea = get_moea_data(n_var, n_objectives, delta_finetune, algorithm, int(fn_eval_limit/pop_size), min_max)
n_evals_moea = n_evals_moea[:-1]
igd_moea = igd_moea[:-1]

Expand All @@ -336,91 +323,8 @@ def main_NSGA():
plt.show()


def main_NSGA_base():
args = get_args()
network_structure = get_network_structure(args)
# generate delta
delta = []
delta.append([np.random.randint(0, 100, args.train_test[0]), np.random.randint(0, 10, args.train_test[0])])
delta.append([np.array([np.mean(delta[0][0])]), np.array([np.mean(delta[0][1])])])
dataset, min_max = get_dataset(args, normalize_targets=True, delta=delta)

n_var = args.problem_dim[0]
n_objectives = args.problem_dim[1]
delta_finetune = np.array(delta[1])[:, -1]

pf_true = get_pf(n_var, n_objectives, delta_finetune, min_max)

igd = []
x_size = []
fn_eval_limit = 300
pop_size = 30
n_gen = int(fn_eval_limit / pop_size)
max_pts_num = 5

saved_x = None
saved_y = None

while sum(x_size) < fn_eval_limit:
sol = train_baseline_nn(None, None, args.problem_dim, init=True)
sol.__dict__['args'] = args

algorithm = NSGA2(pop_size=pop_size)

res = minimize(MyProblem(sol=sol, min_max=min_max),
algorithm,
("n_gen", n_gen),
seed=1,
verbose=False)

X = res.X
if len(X) > max_pts_num:
X = X[np.random.choice(X.shape[0], max_pts_num)]

x_size.append(X.shape[0])

X = X.astype(np.float32)
if saved_x is None:
saved_x = X
else:
saved_x = np.concatenate((saved_x, X))
y_true = evaluate(X, delta_finetune, n_objectives, min_max=min_max)

new_y_true = []
for i in range(n_objectives):
new_y_true.append(y_true[:, i])
new_y_true = np.array(new_y_true, dtype=np.float32)
new_y_true = new_y_true.reshape((*new_y_true.shape, 1))

if saved_y is None:
saved_y = new_y_true
else:
saved_y = np.concatenate((saved_y, new_y_true), axis=1)

# sol.test_continue(X, new_y_true)
train_baseline_nn(saved_x, saved_y, args.problem_dim,
init=False, lr=args.update_lr, n_epochs=100)

metric = IGD(pf_true, zero_to_one=True)
igd.append(metric.do(res.F))

pf = evaluate(res.X, delta_finetune, n_objectives, min_max=min_max)
moea_pf, n_evals_moea, igd_moea = get_moea_data(n_var, n_objectives, delta_finetune, algorithm, n_gen, min_max)
n_evals_moea = n_evals_moea[:-1]
igd_moea = igd_moea[:-1]

visualize_pf(pf=pf, label='Sorrogate PF', color='green', scale=[0.5] * 3, pf_true=pf_true)
visualize_pf(pf=moea_pf, label='NSGA-II PF', color='blue', scale=[0.5] * 3, pf_true=pf_true)

func_evals = [max_pts_num * np.arange(len(igd)), n_evals_moea]
igds = [igd, igd_moea]
colors = ['black', 'red']
labels = ["Our Surrogate Model", "NSGA-II"]
visualize_igd(func_evals, igds, colors, labels)
plt.show()


if __name__ == '__main__':
mains = [main, main_sinewave, main_NSGA, main_NSGA_base]
mains[-2]()
# main()
# main_sinewave()
main_NSGA()

0 comments on commit 1bf9968

Please sign in to comment.