在极大的学习率倍率下可以一定程度上复现论文的效果
This commit is contained in:
parent
1aa5d08034
commit
b4431d20ab
165
main.py
165
main.py
|
@ -1,20 +1,20 @@
|
||||||
|
from tqdm import tqdm
|
||||||
|
from sklearn.preprocessing import MinMaxScaler
|
||||||
|
from keras.layers import Dropout
|
||||||
|
from keras.callbacks import TensorBoard, LearningRateScheduler
|
||||||
|
from keras import regularizers
|
||||||
|
import matplotlib.pyplot as plt
|
||||||
|
import tensorflow as tf
|
||||||
import pandas as pd
|
import pandas as pd
|
||||||
import numpy as np
|
import numpy as np
|
||||||
import os
|
import os
|
||||||
import warnings
|
import warnings
|
||||||
warnings.filterwarnings('ignore')
|
warnings.filterwarnings('ignore')
|
||||||
import tensorflow as tf
|
|
||||||
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
|
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
|
||||||
import matplotlib.pyplot as plt
|
|
||||||
from keras import regularizers
|
|
||||||
from keras.callbacks import TensorBoard, LearningRateScheduler
|
|
||||||
from keras.layers import Dropout
|
|
||||||
from sklearn.preprocessing import MinMaxScaler
|
|
||||||
from tqdm import tqdm
|
|
||||||
|
|
||||||
|
|
||||||
def data_read(data_address):
|
def data_read(data_address):
|
||||||
|
|
||||||
df = pd.read_csv(data_address)
|
df = pd.read_csv(data_address)
|
||||||
label_mapping = {label: idx for idx,
|
label_mapping = {label: idx for idx,
|
||||||
label in enumerate(df['Problem'].unique())}
|
label in enumerate(df['Problem'].unique())}
|
||||||
|
@ -29,64 +29,72 @@ def data_read(data_address):
|
||||||
# 转换为时间序列数据格式
|
# 转换为时间序列数据格式
|
||||||
time_steps = 34
|
time_steps = 34
|
||||||
X_series, Y_series = [], []
|
X_series, Y_series = [], []
|
||||||
for i in range(0, len(X) - time_steps):
|
i = 0
|
||||||
|
while i < len(X):
|
||||||
X_series.append(X[i:(i + time_steps)])
|
X_series.append(X[i:(i + time_steps)])
|
||||||
Y_series.append(Y[i + time_steps - 1])
|
Y_series.append(Y[i + time_steps - 1])
|
||||||
|
i += time_steps
|
||||||
|
|
||||||
return np.array(X_series), np.array(Y_series)
|
return np.array(X_series), np.array(Y_series)
|
||||||
|
|
||||||
|
|
||||||
# 编写绘图函数,画出训练集电压数据
|
# 编写绘图函数,画出训练集电压数据
|
||||||
def plant_for_voltage(x_train_new, x_train_orginal, gamma, learning_rate, y_train, different_location):
|
def plant_for_voltage(x_train_new, x_train_orginal, gamma, learning_rate, y_train, different_location):
|
||||||
# 绘制X_train的图形
|
# 绘制X_train的图形
|
||||||
for i in different_location:
|
for i in different_location:
|
||||||
if i % 100 == 0:
|
if i % 100 == 0:
|
||||||
plt.figure()
|
plt.figure()
|
||||||
time_Step = list(range(0, 34))
|
time_Step = list(range(0, 34))
|
||||||
plt.plot(time_Step,
|
plt.plot(time_Step,
|
||||||
x_train_new[i])
|
x_train_new[i])
|
||||||
# 添加标题和标签
|
# 添加标题和标签
|
||||||
plt.title(f'gamma:{gamma},learning_rate:{learning_rate},Y:{y_train[i]}')
|
plt.title(
|
||||||
plt.xlabel('time_step')
|
f'gamma:{gamma},learning_rate:{learning_rate},Y:{y_train[i]}')
|
||||||
plt.ylabel('voltage')
|
plt.xlabel('time_step')
|
||||||
try:
|
plt.ylabel('voltage')
|
||||||
os.makedirs(f'Liu/picture/gamma{gamma} learningrate{learning_rate}')
|
try:
|
||||||
except FileExistsError:
|
os.makedirs(
|
||||||
pass
|
f'Liu/picture/gamma{gamma} learningrate{learning_rate}')
|
||||||
plt.savefig(f'Liu/picture/gamma{gamma} learningrate{learning_rate}/X_train_new_{i}.png')
|
except FileExistsError:
|
||||||
plt.close()
|
pass
|
||||||
plt.clf()
|
plt.savefig(
|
||||||
|
f'Liu/picture/gamma{gamma} learningrate{learning_rate}/X_train_new_{i}.png')
|
||||||
# 画出原始图像
|
plt.close()
|
||||||
plt.figure()
|
plt.clf()
|
||||||
time_Step = list(range(0, 34))
|
|
||||||
plt.plot(time_Step,
|
|
||||||
x_train_orginal[i])
|
|
||||||
# 添加标题和标签
|
|
||||||
plt.title(f'gamma:{gamma},learning_rate:{learning_rate},Y:{y_train[i]}')
|
|
||||||
plt.xlabel('time_step')
|
|
||||||
plt.ylabel('voltage')
|
|
||||||
try:
|
|
||||||
os.makedirs(f'Liu/picture/gamma{gamma} learningrate{learning_rate}')
|
|
||||||
except FileExistsError:
|
|
||||||
pass
|
|
||||||
plt.savefig(f'Liu/picture/gamma{gamma} learningrate{learning_rate}/X_train_{i}.png')
|
|
||||||
plt.close()
|
|
||||||
plt.clf()
|
|
||||||
|
|
||||||
|
# 画出原始图像
|
||||||
|
plt.figure()
|
||||||
|
time_Step = list(range(0, 34))
|
||||||
|
plt.plot(time_Step,
|
||||||
|
x_train_orginal[i])
|
||||||
|
# 添加标题和标签
|
||||||
|
plt.title(
|
||||||
|
f'gamma:{gamma},learning_rate:{learning_rate},Y:{y_train[i]}')
|
||||||
|
plt.xlabel('time_step')
|
||||||
|
plt.ylabel('voltage')
|
||||||
|
try:
|
||||||
|
os.makedirs(
|
||||||
|
f'Liu/picture/gamma{gamma} learningrate{learning_rate}')
|
||||||
|
except FileExistsError:
|
||||||
|
pass
|
||||||
|
plt.savefig(
|
||||||
|
f'Liu/picture/gamma{gamma} learningrate{learning_rate}/X_train_{i}.png')
|
||||||
|
plt.close()
|
||||||
|
plt.clf()
|
||||||
|
|
||||||
|
|
||||||
def if_diff(a, b):
|
def if_diff(a, b):
|
||||||
# 比较两个数组相同位置上的元素是否相等
|
# 比较两个数组相同位置上的元素是否相等
|
||||||
diff = np.where(a != b)
|
diff = np.where(a != b)
|
||||||
|
|
||||||
list_diff = []
|
list_diff = []
|
||||||
|
|
||||||
|
# 打印不同元素的索引及其对应的元素
|
||||||
|
for i in range(len(diff[0])):
|
||||||
|
idx = (diff[0][i], diff[1][i])
|
||||||
|
list_diff.append(idx[0])
|
||||||
|
return list_diff
|
||||||
|
|
||||||
# 打印不同元素的索引及其对应的元素
|
|
||||||
for i in range(len(diff[0])):
|
|
||||||
idx = (diff[0][i], diff[1][i])
|
|
||||||
list_diff.append(idx[0])
|
|
||||||
return list_diff
|
|
||||||
|
|
||||||
X_train, Y_train = data_read(
|
X_train, Y_train = data_read(
|
||||||
'Liu\data\VOLTAGE-QUALITY-CLASSIFICATION-MODEL--main\Voltage Quality.csv')
|
'Liu\data\VOLTAGE-QUALITY-CLASSIFICATION-MODEL--main\Voltage Quality.csv')
|
||||||
|
@ -122,7 +130,7 @@ model = tf.keras.models.Sequential([
|
||||||
tf.keras.layers.SimpleRNN(100),
|
tf.keras.layers.SimpleRNN(100),
|
||||||
Dropout(0.2),
|
Dropout(0.2),
|
||||||
tf.keras.layers.Dense(n_classes, activation='relu',
|
tf.keras.layers.Dense(n_classes, activation='relu',
|
||||||
) # kernel_regularizer=regularizers.l2(0.3)
|
) # kernel_regularizer=regularizers.l2(0.3)
|
||||||
])
|
])
|
||||||
|
|
||||||
# 编译模型
|
# 编译模型
|
||||||
|
@ -132,30 +140,39 @@ model.compile(
|
||||||
metrics=['accuracy'])
|
metrics=['accuracy'])
|
||||||
|
|
||||||
# 定义学习率指数递减的函数
|
# 定义学习率指数递减的函数
|
||||||
|
|
||||||
|
|
||||||
def lr_schedule(epoch):
|
def lr_schedule(epoch):
|
||||||
initial_learning_rate = 0.01
|
initial_learning_rate = 0.01
|
||||||
decay_rate = 0.1
|
decay_rate = 0.1
|
||||||
decay_steps = 250
|
decay_steps = 250
|
||||||
new_learning_rate = initial_learning_rate * decay_rate ** (epoch / decay_steps)
|
new_learning_rate = initial_learning_rate * \
|
||||||
|
decay_rate ** (epoch / decay_steps)
|
||||||
return new_learning_rate
|
return new_learning_rate
|
||||||
|
|
||||||
|
|
||||||
# 定义学习率调度器
|
# 定义学习率调度器
|
||||||
lr_scheduler = LearningRateScheduler(lr_schedule)
|
lr_scheduler = LearningRateScheduler(lr_schedule)
|
||||||
|
|
||||||
|
|
||||||
# Save initial weights
|
|
||||||
initial_weights = model.get_weights()
|
|
||||||
|
|
||||||
# TensorBoard 回调
|
# TensorBoard 回调
|
||||||
log_dir = "logs/fit"
|
log_dir = "logs/fit"
|
||||||
tensorboard_callback = TensorBoard(log_dir=log_dir, histogram_freq=1)
|
tensorboard_callback = TensorBoard(log_dir=log_dir, histogram_freq=1)
|
||||||
|
|
||||||
|
# 训练模型,添加 TensorBoard 回调
|
||||||
|
model.fit(X_train, Y_train, epochs=500,
|
||||||
|
batch_size=32, callbacks=[tensorboard_callback, lr_scheduler])
|
||||||
|
|
||||||
|
# 评估模型
|
||||||
|
loss, accuracy = model.evaluate(X_test, Y_test)
|
||||||
|
print("Test original accuracy:", accuracy)
|
||||||
|
|
||||||
|
|
||||||
# 制作扰动数据
|
# 制作扰动数据
|
||||||
|
|
||||||
# 转换X_train和Y_train为TensorFlow张量
|
# 转换X_test和Y_test为TensorFlow张量
|
||||||
X_train_tensor = tf.convert_to_tensor(X_train, dtype=tf.float64)
|
X_train_tensor = tf.convert_to_tensor(X_test, dtype=tf.float64)
|
||||||
Y_train_tensor = tf.convert_to_tensor(Y_train, dtype=tf.int32)
|
Y_train_tensor = tf.convert_to_tensor(Y_test, dtype=tf.int32)
|
||||||
|
|
||||||
# 使用tf.GradientTape来计算梯度
|
# 使用tf.GradientTape来计算梯度
|
||||||
with tf.GradientTape() as tape:
|
with tf.GradientTape() as tape:
|
||||||
|
@ -176,7 +193,7 @@ accuracy_per_gamma = {}
|
||||||
flattened_gradients = tf.reshape(gradients, [-1])
|
flattened_gradients = tf.reshape(gradients, [-1])
|
||||||
|
|
||||||
# 选择最大的γ * |X|个梯度
|
# 选择最大的γ * |X|个梯度
|
||||||
for gamma in tqdm([0.6, 0.7, 0.8, 0.9, 0.99]):
|
for gamma in [0.05, 0.1, 0.2, 0.4]:
|
||||||
num_gradients_to_select = int(
|
num_gradients_to_select = int(
|
||||||
gamma * tf.size(flattened_gradients, out_type=tf.dtypes.float32))
|
gamma * tf.size(flattened_gradients, out_type=tf.dtypes.float32))
|
||||||
top_gradients_indices = tf.argsort(flattened_gradients, direction='DESCENDING')[
|
top_gradients_indices = tf.argsort(flattened_gradients, direction='DESCENDING')[
|
||||||
|
@ -200,33 +217,23 @@ for gamma in tqdm([0.6, 0.7, 0.8, 0.9, 0.99]):
|
||||||
# 创建准确率列表
|
# 创建准确率列表
|
||||||
accuracy_list = []
|
accuracy_list = []
|
||||||
|
|
||||||
for learning_rate in tqdm([0.1, 0.2, 0.3, 0.4, 0.5]):
|
for learning_rate in [0.1, 0.2, 0.3, 0.4, 0.5]:
|
||||||
|
|
||||||
# 应用学习率到梯度
|
# 应用学习率到梯度
|
||||||
scaled_gradients = (learning_rate * 100) * updated_gradients
|
scaled_gradients = (learning_rate * 17000) * updated_gradients
|
||||||
# 使用缩放后的梯度更新X_train_tensor
|
# 使用缩放后的梯度更新X_train_tensor
|
||||||
X_train_updated = tf.add(X_train_tensor, scaled_gradients)
|
X_train_updated = tf.add(X_train_tensor, scaled_gradients)
|
||||||
X_train_updated = X_train_updated.numpy()
|
X_train_updated = X_train_updated.numpy()
|
||||||
|
|
||||||
list_diff = if_diff(X_train, X_train_updated)
|
|
||||||
|
|
||||||
# 显示扰动数据和原始数据的可视化图像
|
# 显示扰动数据和原始数据的可视化图像
|
||||||
# plant_for_voltage(X_train_updated, X_train, gamma, learning_rate, Y_train, list_diff)
|
# plant_for_voltage(X_train_updated, X_train, gamma, learning_rate, Y_train, list_diff)
|
||||||
|
|
||||||
# Reset model weights to initial weights
|
|
||||||
model.set_weights(initial_weights)
|
|
||||||
|
|
||||||
# 训练模型,添加 TensorBoard 回调
|
|
||||||
history = model.fit(X_train_updated, Y_train, epochs=500,
|
|
||||||
batch_size=32, callbacks=[tensorboard_callback, lr_scheduler], verbose=0)
|
|
||||||
|
|
||||||
# 评估模型
|
# 评估模型
|
||||||
loss, accuracy = model.evaluate(X_test, Y_test)
|
loss, accuracy = model.evaluate(X_train_updated, Y_test)
|
||||||
print(f"Accuracy gamma: {gamma},learning:{learning_rate}", accuracy)
|
print(f"Accuracy gamma: {gamma},learning:{learning_rate}", accuracy)
|
||||||
|
|
||||||
# 记录准确率
|
# 记录准确率
|
||||||
accuracy_list.append(accuracy)
|
accuracy_list.append(accuracy)
|
||||||
|
|
||||||
|
|
||||||
# 记录该gamma下的准确率
|
# 记录该gamma下的准确率
|
||||||
accuracy_per_gamma[gamma] = accuracy_list
|
accuracy_per_gamma[gamma] = accuracy_list
|
||||||
|
@ -254,5 +261,3 @@ plt.legend()
|
||||||
|
|
||||||
# 显示图像
|
# 显示图像
|
||||||
plt.show()
|
plt.show()
|
||||||
|
|
||||||
# tensorboard --logdir logs/fit
|
|
Loading…
Reference in New Issue