整体思路
- 首先,写一个deep_learning.py文件进行神经网络的训练及测试过程。
- 将deep_learning.py中需要优化的参数(在这里我们优化LSTM层数和全连接层数及每层的神经元个数)统一写到一个列表num中。
- 然后,进行遗传算法的编写GA.py,用需要传入deep_learning.py文件的列表num当染色体,需要优化的参数当染色体上的基因。
deep_learning.py文件
由于要将所有需要优化的参数写到一个列表中,所以再此文件中需要定义两个函数,分别是创建LSTM层函数 create_lstm(inputs, units, return_sequences) 和创建全连接层(包括BN层和dropout层)函数 create_dense(inputs, units) 。
函数:create_lstm(inputs, units, return_sequences)
输入:
- inputs:传进此LSTM层的输入,如果这一LSTM层是第一层LSTM层,则传入的是 layers.Input() 的变量名;否则,传入的应该是上一个LSTM层。
- units:此LSTM层中有多少个神经元。
- return_sequences:此LSTM层保留所有输出(True)还是只保留最后一步的输出(False)。
输出:
- 输出LSTM层。
# 定义LSTM层函数
def create_lstm(inputs, units, return_sequences):
lstm = layers.Bidirectional(layers.LSTM(units, return_sequences=return_sequences))(inputs)
print('Lstm', lstm.shape)
return lstm
函数:create_dense(inputs, units)
输入:
- inputs:传进此全连接层的输入,如果这一全连接层是第一层全连接层,则传入的是 layers.Flatten() 的变量名;否则,传入的应该是上一个全连接层。
- units:此全连接层中有多少个神经元。
输出:
- 输出全连接层、BN层和dropout层。
# 定义Dense层函数
def create_dense(inputs, units):
dense = layers.Dense(units, kernel_regularizer=keras.regularizers.l2(0.001), activation='relu')(inputs)
print('Dense', dense.shape)
dense_dropout = layers.Dropout(0.2)(dense)
dense_batch = layers.BatchNormalization()(dense_dropout)
return dense, dense_dropout, dense_batch
设置参数
设置LSTM层参数的时候,只有最后一层只保留最后一步的输出,其他的都是全部保留。
# 设置LSTM层参数
lstm_num_layers = 2
lstm_units = [128, 128]
lstm_name = list(np.zeros((lstm_num_layers,)))
# 设置LSTM_Dense层参数
lstm_dense_num_layers = 2
lstm_dense_units = [128, 64]
lstm_dense_name = list(np.zeros((lstm_dense_num_layers,)))
lstm_dense_dropout_name = list(np.zeros((lstm_dense_num_layers,)))
lstm_dense_batch_name = list(np.zeros((lstm_dense_num_layers,)))
调用函数构建模型
按照介绍函数时的解释构建网络模型。
inputs_lstm = layers.Input(shape=(x_train.shape[1], x_train.shape[2]))
print(inputs_lstm.shape)
for i in range(lstm_num_layers):
if i == 0:
inputs = inputs_lstm
else:
inputs = lstm_name[i-1]
if i == lstm_num_layers - 1:
return_sequences=False
else:
return_sequences=True
lstm_name[i] = create_lstm(inputs, lstm_units[i], return_sequences)
for i in range(lstm_dense_num_layers):
if i == 0:
inputs = lstm_name[lstm_num_layers-1]
else:
inputs = lstm_dense_batch_name[i-1]
lstm_dense_name[i], lstm_dense_dropout_name[i], lstm_dense_batch_name[i] = create_dense(inputs, lstm_dense_units[i])
outputs_lstm = layers.Dense(10, activation='softmax')(lstm_dense_batch_name[lstm_dense_num_layers-1])
print('Outputs:', outputs_lstm.shape)
完整代码
以上没有用到列表num,而是直接将层数设为2,神经元数量也直接给出,目的是为了方便讲解,下面给出完整代码:
import tensorflow as tf
import tensorflow.keras as keras
from tensorflow.keras import models, layers, optimizers
import matplotlib.pyplot as plt
# 定义LSTM层函数
def create_lstm(inputs, units, return_sequences):
lstm = layers.Bidirectional(layers.LSTM(units, return_sequences=return_sequences))(inputs)
return lstm
# 定义Dense层函数
def create_dense(inputs, units):
dense = layers.Dense(units, kernel_regularizer=keras.regularizers.l2(0.001), activation='relu')(inputs)
dense_dropout = layers.Dropout(0.2)(dense)
dense_batch = layers.BatchNormalization()(dense_dropout)
return dense, dense_dropout, dense_batch
def load():
# Mnist数据集加载
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
# Mnist数据集简单归一化
x_train, x_test = x_train / 255.0, x_test / 255.0
return x_train, y_train, x_test, y_test
def classify(x_train, y_train, x_test, y_test, num):
# 设置LSTM层参数
lstm_num_layers = num[0]
lstm_units = num[2: 2+lstm_num_layers]
lstm_name = list(np.zeros((lstm_num_layers,)))
# 设置LSTM_Dense层参数
lstm_dense_num_layers = num[1]
lstm_dense_units = num[2+lstm_num_layers: 2+lstm_num_layers+lstm_dense_num_layers]
lstm_dense_name = list(np.zeros((lstm_dense_num_layers,)))
lstm_dense_dropout_name = list(np.zeros((lstm_dense_num_layers,)))
lstm_dense_batch_name = list(np.zeros((lstm_dense_num_layers,)))
inputs_lstm = layers.Input(shape=(x_train.shape[1], x_train.shape[2]))
for i in range(lstm_num_layers):
if i == 0:
inputs = inputs_lstm
else:
inputs = lstm_name[i-1]
if i == lstm_num_layers - 1:
return_sequences=False
else:
return_sequences=True
lstm_name[i] = create_lstm(inputs, lstm_units[i], return_sequences)
for i in range(lstm_dense_num_layers):
if i == 0:
inputs = lstm_name[lstm_num_layers-1]
else:
inputs = lstm_dense_batch_name[i-1]
lstm_dense_name[i], lstm_dense_dropout_name[i], lstm_dense_batch_name[i] = create_dense(inputs, lstm_dense_units[i])
outputs_lstm = layers.Dense(10, activation='softmax')(lstm_dense_batch_name[lstm_dense_num_layers-1])
LSTM_model = keras.Model(inputs_lstm, outputs_lstm)
LSTM_model.compile(optimizer=keras.optimizers.Adam(),
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
history = LSTM_model.fit(x_train, y_train, batch_size=32, epochs=5, validation_split=0.1, verbose=0)
# 验证模型:
lstm.evaluate(x_test, y_test, verbose=0)
return results[1]
列表num中的前两个元素分别表示LSTM层的层数和全连接层的层数,后面的元素表示每层的神经元个数。
返回的值为测试集的准确率。
GA.py
常规的遗传算法介绍可以参考我的另一篇文章遗传算法求解最大值问题详解(附python代码)。
问题
在优化卷积神经网络这个问题上,用常规的遗传算法不易实现,原因如下:
- 1、传统的遗传算法中的每条染色体的长度相同,但是优化LSTM网络时,染色体的长度会因为层数的不同而不同。比如a染色体有一层LSTM层和一层全连接层,则在这条染色体上共有四个基因(两个代表层数,两个代表每层的神经元个数);b染色体有两层LSTM层和两层全连接层,则在这条染色体上共有六个基因(两个代表层数,四个代表每层的神经元个数)。
- 2、在传统的遗传算法中,染色体上的基因的取值范围都是相同的,但优化LSTM网络时,需要让表示层数的基因在一个范围内,表示神经元个数的基因在另一个范围内。比如,LSTM层层数在一层到三层之间,全连接层个数在一层到三层之间,神经元个数在32个到256个之间。
- 3、由于第一个问题(即染色体长度不同)的存在,交叉函数、变异函数均需要做出修改。
解决方法
- 1、将每条染色体设置为相同的长度(因为LSTM层层数最多三层,全连接层个数最多三层,加上最前面两个表示层数的基因,故设置每条染色体上有3+3+2=8个基因),达不到长度要求的后面补零。
- 2、先设置前面两个基因,令其范围分别在一到三之间和一到三之间,然后根据这连个基因确定后面关于每层神经元个数的基因的个数。
- 3、对于交叉函数的修改,首先确定取出的两条染色体(设为a染色体和b染色体)上需要交换的位置,然后遍历两条染色体在这些位置上的基因,如果任一染色体上此位置上的基因为0或要交换的基因是关于层数的,则取消此位置的交换。
- 4、对于变异函数的修改,只有关于神经元个数的基因变异,关于层数的基因不变异。
完整代码
import numpy as np
import deep_learning as project
DNA_SIZE = 2
DNA_SIZE_MAX = 8
POP_SIZE = 20
CROSS_RATE = 0.5
MUTATION_RATE = 0.01
N_GENERATIONS = 40
train_x, train_y, test_x, test_y = project.load()
def get_fitness(x):
return project.classify(train_x, train_y, test_x, test_y, num=x)
def select(pop, fitness):
idx = np.random.choice(np.arange(POP_SIZE), size=POP_SIZE, replace=True, p=fitness / fitness.sum())
return pop[idx]
def crossover(parent, pop):
if np.random.rand() < CROSS_RATE:
i_ = np.random.randint(0, POP_SIZE, size=1)
cross_points = np.random.randint(0, 2, size=DNA_SIZE_MAX).astype(np.bool)
for i, point in enumerate(cross_points):
if point == True and pop[i_, i]*parent[i] == 0:
cross_points[i] = False
if point == True and i < 2:
cross_points[i] = False
parent[cross_points] = pop[i_, cross_points]
return parent
def mutate(child):
for point in range(DNA_SIZE_MAX):
if np.random.rand() < MUTATION_RATE:
if point >= 3:
if child[point] != 0:
child[point] = np.random.randint(32, 512)
return child
pop_layers = np.zeros((POP_SIZE, DNA_SIZE), np.int32)
pop_layers[:, 0] = np.random.randint(1, 4, size=(POP_SIZE,))
pop_layers[:, 1] = np.random.randint(1, 4, size=(POP_SIZE,))
pop = np.zeros((POP_SIZE, DNA_SIZE_MAX))
for i in range(POP_SIZE):
pop_neurons = np.random.randint(32, 257, size=(pop_layers[i].sum(),))
pop_stack = np.hstack((pop_layers[i], pop_neurons))
for j, gene in enumerate(pop_stack):
pop[i][j] = gene
for each_generation in range(N_GENERATIONS):
fitness = np.zeros([POP_SIZE, ])
for i in range(POP_SIZE):
pop_list = list(pop[i])
for j, each in enumerate(pop_list):
if each == 0.0:
index = j
pop_list = pop_list[:j]
for k, each in enumerate(pop_list):
each_int = int(each)
pop_list[k] = each_int
fitness[i] = get_fitness(pop_list)
print('第%d代第%d个染色体的适应度为%f' % (each_generation+1, i+1, fitness[i]))
print('此染色体为:', pop_list)
print("Generation:", each_generation+1, "Most fitted DNA: ", pop[np.argmax(fitness), :], "适应度为:", fitness[np.argmax(fitness)])
pop = select(pop, fitness)
pop_copy = pop.copy()
for parent in pop:
child = crossover(parent, pop_copy)
child = mutate(child)
parent = child
其中,如下代码的作用是将数组中的0元素删除掉,具体实现过程可以参考我的另一篇文章删掉nd array数组中的所有零元素。
for each_generation in range(N_GENERATIONS):
fitness = np.zeros([POP_SIZE, ])
for i in range(POP_SIZE):
pop_list = list(pop[i])
for j, each in enumerate(pop_list):
if each == 0.0:
index = j
pop_list = pop_list[:j]
for k, each in enumerate(pop_list):
each_int = int(each)
pop_list[k] = each_int