jung-geun/PSO
1
0
Fork 0
mirror of https://github.com/jung-geun/PSO.git synced 2025-12-19 20:44:39 +09:00
Code Issues Packages Projects Releases Wiki Activity

chore: 업데이트된 패키지 요구사항 반영

Browse Source 
requirements.txt 파일에서 패키지 요구사항을 업데이트했습니다.
This commit is contained in:
jung-geun
2024-05-16 02:11:31 +09:00
parent 4a6a48d6aa
commit 737baf6681
4 changed files with 183 additions and 325 deletions
Download Patch File Download Diff File Expand all files Collapse all files

245
pso/optimizer.py
View File

@@ -49,7 +49,7 @@ class Optimizer:
w_min (float): 최소 관성 수치
w_max (float): 최대 관성 수치
negative_swarm (float): 최적해와 반대로 이동할 파티클 비율 - 0 ~ 1 사이의 값
mutation_swarm (float): 돌연변이가 일어날 확률
mutation_swarm (float): 돌연변이가 일어날 확률 - 0 ~ 1 사이의 값
np_seed (int | None): numpy seed. Defaults to None.
tf_seed (int | None): tensorflow seed. Defaults to None.
random_state (tuple): numpy random state. Defaults to None.
@@ -105,6 +105,7 @@ class Optimizer:
model.compile(loss=loss, optimizer="adam", metrics=["accuracy", "mse"])
self.model = model # 모델 구조
self.set_shape(model.get_weights())
self.loss = loss # 손실함수
self.n_particles = n_particles # 파티클 개수
self.particles = [None] * n_particles # 파티클 리스트
@@ -187,7 +188,19 @@ class Optimizer:
tf.keras.backend.reset_uids()
tf.keras.backend.clear_session()
def _encode(self, weights):
def set_shape(self, weights: list):
"""
가중치의 shape을 설정
Args:
weights (list): keras model의 가중치
"""
self.shape = [layer.shape for layer in weights]
def get_shape(self):
return self.shape
def _encode(self, weights: list):
"""
가중치를 1차원으로 풀어서 반환
@@ -199,19 +212,13 @@ class Optimizer:
(list) : 가중치의 원본 shape의 길이
"""
w_gpu = np.array([])
length = []
shape = []
for layer in weights:
shape.append(layer.shape)
w_tmp = layer.reshape(-1)
length.append(len(w_tmp))
w_gpu = np.append(w_gpu, w_tmp)
del weights
return w_gpu
return w_gpu, shape, length
def _decode_(self, weight, shape, length):
def _decode(self, weight: np.ndarray):
"""
_encode 로 인코딩된 가중치를 원본 shape으로 복원
파라미터는 encode의 리턴값을 그대로 사용을 권장
@@ -225,15 +232,15 @@ class Optimizer:
"""
weights = []
start = 0
for i in range(len(shape)):
end = start + length[i]
w_tmp = weight[start:end]
w_tmp = np.reshape(w_tmp, shape[i])
weights.append(w_tmp)
for i in range(len(self.shape)):
end = start + np.prod(self.shape[i])
w_ = weight[start:end]
w_ = np.reshape(w_, self.shape[i])
weights.append(w_)
start = end
del weight, shape, length
del start, end, w_tmp
del start, end, w_
del weight
return weights
@@ -271,11 +278,12 @@ class Optimizer:
(float): 가중치의 최소값
(float): 가중치의 최대값
"""
w_, w_s, w_l = self._encode(Particle.g_best_weights)
w_ = self._encode(self.model.get_weights())
# w_, w_s, w_l = self._encode(Particle.g_best_weights)
weight_min = np.min(w_)
weight_max = np.max(w_)
del w_, w_s, w_l
del w_
return weight_min, weight_max
@@ -290,9 +298,12 @@ class Optimizer:
self.index += 1
if self.index > self.max_index:
self.index = 0
self.__get_batch_slice(self.batch_size)
self.dataset = self.__get_batch_slice(self.batch_size)
return self.dataset[self.index - 1][0], self.dataset[self.index - 1][1]
def get_length(self):
return self.get_max_index()
def get_max_index(self):
return self.max_index
@@ -342,8 +353,6 @@ class Optimizer:
save_info : bool - 종료시 학습 정보 저장 여부 default : False,
save_path : str - ex) "./result",
renewal : str ex) "acc" or "loss" or "mse",
empirical_balance : bool - True : EBPSO, False : PSO,
dispersion : bool - True : g_best 의 값을 분산시켜 전역해를 찾음, False : g_best 의 값만 사용
check_point : int - 저장할 위치 - None : 저장 안함
batch_size : int - batch size default : None => len(x) // 10
batch_size > len(x) : auto max batch size
@@ -357,8 +366,6 @@ class Optimizer:
log_name = kwargs.get("log_name", None)
save_info = kwargs.get("save_info", False)
renewal = kwargs.get("renewal", "acc")
empirical_balance = kwargs.get("empirical_balance", False)
dispersion = kwargs.get("dispersion", False)
check_point = kwargs.get("check_point", None)
batch_size = kwargs.get("batch_size", None)
validate_data = kwargs.get("validate_data", None)
@@ -387,12 +394,6 @@ class Optimizer:
elif renewal not in ["acc", "loss", "mse"]:
raise ValueError("renewal not in ['acc', 'loss', 'mse']")
if empirical_balance is None:
empirical_balance = False
if dispersion is None:
dispersion = False
if (
validate_data is not None
and validate_data[0].shape[0] != validate_data[1].shape[0]
@@ -425,9 +426,6 @@ class Optimizer:
print(e)
sys.exit(10)
self.empirical_balance = empirical_balance
self.dispersion = dispersion
self.renewal = renewal
try:
@@ -484,19 +482,7 @@ class Optimizer:
del model_
else:
for i in tqdm(
range(len(self.particles)),
desc="best score init",
ascii=True,
leave=False,
):
score = self.particles[i].get_score(
validate_data[0], validate_data[1], self.renewal
)
self.particles[i].check_global_best(self.renewal)
print("best score init complete" + str(Particle.g_best_score))
print("best score init complete" + str(Particle.g_best_score))
epochs_pbar = tqdm(
range(epochs),
@@ -505,6 +491,7 @@ class Optimizer:
leave=True,
position=0,
)
rng = np.random.default_rng(seed=42)
for epoch in epochs_pbar:
# 이번 epoch의 평균 점수
# particle_avg = particle_sum / self.n_particles # x_j
@@ -530,117 +517,67 @@ class Optimizer:
* (epoch % weight_reduction)
/ weight_reduction
)
rng = np.random.default_rng(seed=42)
for i in part_pbar:
part_pbar.set_description(
f"loss: {min_loss:.4f} acc: {max_acc:.4f} mse: {min_mse:.4f}"
)
g_best = Particle.g_best_weights
x_batch, y_batch = dataset.next()
weight_min, weight_max = self.__weight_range()
if dispersion:
ts = weight_min + rng.random() * (weight_max - weight_min)
g_, g_sh, g_len = self._encode(Particle.g_best_weights)
decrement = (epochs - epoch + 1) / epochs
g_ = (1 - decrement) * g_ + decrement * ts
g_best = self._decode_(g_, g_sh, g_len)
if empirical_balance:
if rng.random() < np.exp(-(epoch) / epochs):
w_p_ = self._f(
x_batch, y_batch, self.particles[i].get_best_weights()
)
w_g_ = self._f(x_batch, y_batch, g_best)
w_p = w_p_ / (w_p_ + w_g_)
w_g = w_p_ / (w_p_ + w_g_)
del w_p_
del w_g_
else:
p_b = self.particles[i].get_best_score()
g_a = self.avg_score
l_b = p_b[1] - g_a
sigma_post = np.sqrt(np.power(l_b, 2))
sigma_pre = (
1
/ (
self.n_particles
* np.linalg.norm(weight_min - weight_max)
)
* sigma_post
)
p_ = np.exp(-1 * sigma_pre * sigma_post)
w_p = p_
w_g = 1 - p_
del p_b
del g_a
del l_b
del p_
score = self.particles[i].step_w(
x_batch,
y_batch,
self.c0,
self.c1,
w,
w_p,
w_g,
renewal=renewal,
for _i in tqdm(
range(dataset.get_length()),
desc="batch",
ascii=True,
leave=False,
):
part_pbar.set_description(
f"loss: {min_loss:.4f} acc: {max_acc:.4f} mse: {min_mse:.4f}"
)
x_batch, y_batch = dataset.next()
else:
score = self.particles[i].step(
x_batch, y_batch, self.c0, self.c1, w, renewal=renewal
)
if renewal == "loss":
# 최저 loss 보다 작거나 같을 경우
if score[0] < min_loss:
# 각 점수 갱신
min_loss, max_acc, min_mse = score
best_particle_index = i
elif score[0] == min_loss:
if score[1] > max_acc:
min_loss, max_acc, min_mse = score
best_particle_index = i
elif renewal == "acc":
# 최고 점수 보다 높거나 같을 경우
if score[1] > max_acc:
# 각 점수 갱신
min_loss, max_acc, min_mse = score
best_particle_index = i
elif score[1] == max_acc:
if score[2] < min_mse:
min_loss, max_acc, min_mse = score
best_particle_index = i
elif renewal == "mse":
if score[2] < min_mse:
min_loss, max_acc, min_mse = score
best_particle_index = i
elif score[2] == min_mse:
if score[1] > max_acc:
min_loss, max_acc, min_mse = score
best_particle_index = i
if log == 2:
with self.train_summary_writer[i].as_default():
tf.summary.scalar("accuracy", score[1], step=epoch + 1)
tf.summary.scalar("loss", score[0], step=epoch + 1)
tf.summary.scalar("mse", score[2], step=epoch + 1)
if renewal == "loss":
# 최저 loss 보다 작거나 같을 경우
if score[0] < min_loss:
# 각 점수 갱신
min_loss, max_acc, min_mse = score
best_particle_index = i
elif score[0] == min_loss:
if score[1] > max_acc:
min_loss, max_acc, min_mse = score
best_particle_index = i
elif renewal == "acc":
# 최고 점수 보다 높거나 같을 경우
if score[1] > max_acc:
# 각 점수 갱신
min_loss, max_acc, min_mse = score
best_particle_index = i
elif score[1] == max_acc:
if score[2] < min_mse:
min_loss, max_acc, min_mse = score
best_particle_index = i
elif renewal == "mse":
if score[2] < min_mse:
min_loss, max_acc, min_mse = score
best_particle_index = i
elif score[2] == min_mse:
if score[1] > max_acc:
min_loss, max_acc, min_mse = score
best_particle_index = i
if log == 1:
with open(
f"./logs/{log_name}/{self.day}/{self.n_particles}_{epochs}_{self.c0}_{self.c1}_{self.w_min}_{renewal}.csv",
@@ -651,6 +588,7 @@ class Optimizer:
f.write(", ")
else:
f.write("\n")
part_pbar.refresh()
# 한번 epoch 가 끝나고 갱신을 진행해야 순간적으로 높은 파티클이 발생해도 오류가 생기지 않음
if renewal == "loss" and min_loss <= Particle.g_best_score[0]:
@@ -717,14 +655,17 @@ class Optimizer:
(keras.models): 모델
"""
model = keras.models.model_from_json(self.model.to_json())
model.set_weights(Particle.g_best_weights)
model.compile(
loss=self.loss,
optimizer="adam",
metrics=["accuracy", "mse"],
)
if Particle.g_best_weights is not None:
model.set_weights(self._decode(Particle.g_best_weights))
model.compile(
loss=self.loss,
optimizer="adam",
metrics=["accuracy", "mse"],
)
return model
return model
else:
return None
def get_best_score(self):
"""
@@ -800,6 +741,10 @@ class Optimizer:
"""
x, y = valid_data
model = self.get_best_model()
if model is None:
return None
score = model.evaluate(x, y, verbose=1) # type: ignore
print(f"model score - loss: {score[0]} - acc: {score[1]} - mse: {score[2]}")

235
pso/particle.py
View File

@@ -1,6 +1,7 @@
from typing import Any
import numpy as np
from tensorflow import keras
from typing import Any
class Particle:
@@ -40,7 +41,8 @@ class Particle:
converge_reset (bool, optional): 조기 종료 사용 여부. Defaults to False.
converge_reset_patience (int, optional): 조기 종료를 위한 기다리는 횟수. Defaults to 10.
"""
self.model = model
self.set_model(model)
self.weights = self._encode(model.get_weights())
self.loss = loss
try:
@@ -61,11 +63,12 @@ class Particle:
print(e)
exit(1)
self.velocities = np.zeros(len(self.weights))
self.__reset_particle()
self.best_weights = self.get_weights()
self.best_weights = self.weights
self.negative = negative
self.mutation = mutation
self.best_score = [np.inf, 0, np.inf]
self.local_best_score = [np.inf, 0, np.inf]
self.score_history = []
self.converge_reset = converge_reset
self.converge_reset_patience = converge_reset_patience
@@ -78,10 +81,22 @@ class Particle:
del self.loss
del self.velocities
del self.negative
del self.best_score
del self.local_best_score
del self.best_weights
Particle.count -= 1
def set_shape(self, weights: list):
"""
가중치의 shape을 설정
Args:
weights (list): keras model의 가중치
"""
self.shape = [layer.shape for layer in weights]
def get_shape(self):
return self.shape
def _encode(self, weights: list):
"""
가중치를 1차원으로 풀어서 반환
@@ -94,17 +109,13 @@ class Particle:
(list) : 가중치의 원본 shape의 길이
"""
w_gpu = np.array([])
length = []
shape = []
for layer in weights:
shape.append(layer.shape)
w_tmp = layer.reshape(-1)
length.append(len(w_tmp))
w_gpu = np.append(w_gpu, w_tmp)
return w_gpu, shape, length
return w_gpu
def _decode(self, weight: list, shape, length):
def _decode(self, weight: np.ndarray):
"""
_encode 로 인코딩된 가중치를 원본 shape으로 복원
파라미터는 encode의 리턴값을 그대로 사용을 권장
@@ -118,15 +129,14 @@ class Particle:
"""
weights = []
start = 0
for i in range(len(shape)):
end = start + length[i]
for i in range(len(self.shape)):
end = start + np.prod(self.shape[i])
w_ = weight[start:end]
w_ = np.reshape(w_, shape[i])
w_ = np.reshape(w_, self.shape[i])
weights.append(w_)
start = end
del start, end, w_
del shape, length
del weight
return weights
@@ -139,6 +149,7 @@ class Particle:
def set_model(self, model: keras.Model):
self.model = model
self.set_shape(self.model.get_weights())
def compile(self):
if self.model is None:
@@ -151,10 +162,9 @@ class Particle:
)
def get_weights(self):
if self.model is None:
raise ValueError(self.MODEL_IS_NONE)
weights = self._decode(self.weights)
return self.model.get_weights()
return weights
def evaluate(self, x, y):
if self.model is None:
@@ -177,17 +187,17 @@ class Particle:
score = self.evaluate(x, y)
if renewal == "loss":
if score[0] < self.best_score[0]:
self.best_score = score
self.best_weights = self.get_weights()
if score[0] < self.local_best_score[0]:
self.local_best_score = score
self.best_weights = self.weights
elif renewal == "acc":
if score[1] > self.best_score[1]:
self.best_score = score
self.best_weights = self.get_weights()
if score[1] > self.local_best_score[1]:
self.local_best_score = score
self.best_weights = self.weights
elif renewal == "mse":
if score[2] < self.best_score[2]:
self.best_score = score
self.best_weights = self.get_weights()
if score[2] < self.local_best_score[2]:
self.local_best_score = score
self.best_weights = self.weights
else:
raise ValueError("renewal must be 'acc' or 'loss' or 'mse'")
@@ -234,12 +244,10 @@ class Particle:
loss=self.loss,
metrics=["accuracy", "mse"],
)
i_w_, i_s, i_l = self._encode(self.get_weights())
self.weights = self._encode(self.model.get_weights())
rng = np.random.default_rng()
i_w_ = rng.uniform(-0.1, 0.1, len(i_w_))
self.velocities = self._decode(i_w_, i_s, i_l)
self.velocities = rng.uniform(-0.2, 0.2, len(self.weights))
del i_w_, i_s, i_l
self.score_history = []
def _velocity_calculation(self, local_rate, global_rate, w):
@@ -251,10 +259,12 @@ class Particle:
global_rate (float): 전역 최적해의 영향력
w (float): 현재 속도의 영향력 - 관성 | 0.9 ~ 0.4 이 적당
"""
encode_w, w_sh, w_len = self._encode(weights=self.get_weights())
encode_v, v_sh, v_len = self._encode(weights=self.velocities)
encode_p, p_sh, p_len = self._encode(weights=self.best_weights)
encode_g, g_sh, g_len = self._encode(weights=Particle.g_best_weights)
# 0회차 전역 최적해가 없을 경우 현재 파티클의 최적해로 설정 - 전역최적해의 방향을 0으로 만들기 위함
best_particle_weights = (
self.best_weights
if Particle.g_best_weights is None
else Particle.g_best_weights
)
rng = np.random.default_rng(seed=42)
r_0 = rng.random()
@@ -263,9 +273,9 @@ class Particle:
if self.negative:
# 지역 최적해와 전역 최적해를 음수로 사용하여 전역 탐색을 유도
new_v = (
w * encode_v
+ local_rate * r_0 * (encode_p - encode_w)
- global_rate * r_1 * (encode_g - encode_w)
w * self.velocities
+ local_rate * r_0 * (self.best_weights - self.weights)
- global_rate * r_1 * (best_particle_weights - self.weights)
)
if (
len(self.score_history) > 10
@@ -274,81 +284,35 @@ class Particle:
self.__reset_particle()
else:
# 전역 최적해의 acc 가 높을수록 더 빠르게 수렴
# 하지만 loss 가 커진 상태에서는 전역 최적해의 영향이
new_v = (
w * encode_v
+ local_rate * r_0 * (encode_p - encode_w)
+ global_rate * r_1 * (encode_g - encode_w)
w * self.velocities
+ local_rate
* self.local_best_score[1]
* r_0
* (self.best_weights - self.weights)
+ global_rate
* Particle.g_best_score[1]
* r_1
* (best_particle_weights - self.weights)
)
if rng.random() < self.mutation:
m_v = rng.uniform(-0.1, 0.1, len(encode_v))
if self.mutation != 0.0 and rng.random() < self.mutation:
m_v = rng.uniform(-0.2, 0.2, len(self.velocities))
new_v = m_v
self.velocities = self._decode(new_v, w_sh, w_len)
self.velocities = new_v
del encode_w, w_sh, w_len
del encode_v, v_sh, v_len
del encode_p, p_sh, p_len
del encode_g, g_sh, g_len
del r_0, r_1
def _velocity_calculation_w(self, local_rate, global_rate, w, w_p, w_g):
"""
현재 속도 업데이트
기본 업데이트의 변형으로 지역 최적해와 전역 최적해를 분산시켜 조기 수렴을 방지
Args:
local_rate (float): 지역 최적해의 영향력
global_rate (float): 전역 최적해의 영향력
w (float): 현재 속도의 영향력 - 관성 | 0.9 ~ 0.4 이 적당
w_p (float): 지역 최적해의 분산 정도
w_g (float): 전역 최적해의 분산 정도
"""
encode_w, w_sh, w_len = self._encode(weights=self.get_weights())
encode_v, v_sh, v_len = self._encode(weights=self.velocities)
encode_p, p_sh, p_len = self._encode(weights=self.best_weights)
encode_g, g_sh, g_len = self._encode(weights=Particle.g_best_weights)
rng = np.random.default_rng(seed=42)
r_0 = rng.random()
r_1 = rng.random()
if self.negative:
new_v = (
w * encode_v
+ local_rate * r_0 * (w_p * encode_p - encode_w)
- global_rate * r_1 * (w_g * encode_g - encode_w)
)
else:
new_v = (
w * encode_v
+ local_rate * r_0 * (w_p * encode_p - encode_w)
+ global_rate * r_1 * (w_g * encode_g - encode_w)
)
if rng.random() < self.mutation:
m_v = rng.uniform(-0.1, 0.1, len(encode_v))
new_v = m_v
self.velocities = self._decode(new_v, w_sh, w_len)
del encode_w, w_sh, w_len
del encode_v, v_sh, v_len
del encode_p, p_sh, p_len
del encode_g, g_sh, g_len
del r_0, r_1
def _position_update(self):
"""
가중치 업데이트
"""
encode_w, w_sh, w_len = self._encode(weights=self.get_weights())
encode_v, v_sh, v_len = self._encode(weights=self.velocities)
new_w = encode_w + encode_v
self.model.set_weights(self._decode(new_w, w_sh, w_len))
new_w = np.add(self.weights, self.velocities)
del encode_w, w_sh, w_len
del encode_v, v_sh, v_len
self.model.set_weights(self._decode(new_w))
def step(self, x, y, local_rate, global_rate, w, renewal: str = "acc"):
"""
@@ -399,58 +363,6 @@ class Particle:
return score
def step_w(self, x, y, local_rate, global_rate, w, w_p, w_g, renewal: str = "acc"):
"""
파티클의 한 스텝을 진행합니다.
기본 스텝의 변형으로, 지역최적해와 전역최적해의 분산 정도를 조정할 수 있습니다
Args:
x (list): 입력 데이터
y (list): 출력 데이터
local_rate (float): 지역 최적해의 영향력
global_rate (float): 전역 최적해의 영향력
w (float): 관성
g_best (list): 전역 최적해
w_p (float): 지역 최적해의 분산 정도
w_g (float): 전역 최적해의 분산 정도
renewal (str, optional): 최고점수 갱신 방식. Defaults to "acc" | "acc" or "loss"
Returns:
float: 현재 파티클의 점수
"""
self._velocity_calculation_w(local_rate, global_rate, w, w_p, w_g)
self._position_update()
score = self.get_score(x, y, renewal)
if self.converge_reset and self.__check_converge_reset(
score,
self.converge_reset_monitor,
self.converge_reset_patience,
self.converge_reset_min_delta,
):
self.__reset_particle()
score = self.get_score(x, y, renewal)
while (
np.isnan(score[0])
or np.isnan(score[1])
or np.isnan(score[2])
or score[0] == 0
or score[1] == 0
or score[2] == 0
or np.isinf(score[0])
or np.isinf(score[1])
or np.isinf(score[2])
or score[0] > 1000
or score[1] > 1
or score[2] > 1000
):
self.__reset_particle()
score = self.get_score(x, y, renewal)
return score
def get_best_score(self):
"""
파티클의 최고점수를 반환합니다.
@@ -458,7 +370,7 @@ class Particle:
Returns:
float: 최고점수
"""
return self.best_score
return self.local_best_score
def get_best_weights(self):
"""
@@ -467,11 +379,11 @@ class Particle:
Returns:
list: 가중치 리스트
"""
return self.best_weights
return self._decode(self.best_weights)
def set_global_score(self):
"""전역 최고점수를 현재 파티클의 최고점수로 설정합니다"""
Particle.g_best_score = self.best_score
Particle.g_best_score = self.local_best_score
def set_global_weights(self):
"""전역 최고점수를 받은 가중치를 현재 파티클의 최고점수를 받은 가중치로 설정합니다"""
@@ -484,8 +396,15 @@ class Particle:
def check_global_best(self, renewal: str = "loss"):
if (
(renewal == "loss" and self.best_score[0] < Particle.g_best_score[0])
or (renewal == "acc" and self.best_score[1] > Particle.g_best_score[1])
or (renewal == "mse" and self.best_score[2] < Particle.g_best_score[2])
(renewal == "loss" and self.local_best_score[0] < Particle.g_best_score[0])
or (
renewal == "acc" and self.local_best_score[1] > Particle.g_best_score[1]
)
or (
renewal == "mse" and self.local_best_score[2] < Particle.g_best_score[2]
)
):
self.update_global_best()
# 끝

9
requirements.txt
View File

@@ -1,8 +1,7 @@
ipython
keras
numpy
pandas
tensorflow
tqdm
scikit-learn
tensorboard
tensorflow==2.15.1
tqdm==4.66.4
scikit-learn==1.4.2
tensorboard==2.15.1

19
setup.py
View File

@@ -1,9 +1,14 @@
from setuptools import setup, find_packages
from setuptools import find_packages, setup
import pso
VERSION = pso.__version__
def get_requirements(path: str):
return [l.strip() for l in open(path)]
setup(
name="pso2keras",
version=VERSION,
@@ -11,17 +16,7 @@ setup(
author="pieroot",
author_email="jgbong0306@gmail.com",
url="https://github.com/jung-geun/PSO",
install_requires=[
"tqdm",
"numpy",
"pandas",
"ipython",
"matplotlib",
"tensorflow",
"keras",
"scikit-learn",
"tensorboard",
],
install_requires=get_requirements("requirements.txt"),
packages=find_packages(exclude=[]),
keywords=[
"pso",