一. 数据集预处理
1. 加载csv文件
obj = pd.read_csv("文件路径", header=0)
2. 提取数据
data = ([:, m:n+1]) # 数据集中在m~n列
3. 数据归一化
参考数据标准化处理
4. 数据输入格式基本处理
# 如将样本处理成三维数据,可利用reshape()
data = (num, rows, cols)
# 列表转数组
data = (data)
5. 提取标签值
# 提取标签值
obj = obj[['列名1','列名2',...,'列名n']]
6. 标签值的类别转id(即用数字表示标签的类别数,如果有n类标签就用0~n-1表示)
# 假设只有一列为标签
obj['id'] = obj['列名'].factorize()[0] # 在其后增加一列,列名为id
7. one-hot编码(适用于多分类问题)
# 将id列的数值进行编码
''' 如0~5编码后对应[[1,0,0,0,0,0],
[0,1,0,0,0,0],
[0,0,1,0,0,0],
[0,0,0,1,0,0],
[0,0,0,0,1,0],
[0,0,0,0,0,1]]
'''
lable = pd.get_dummies(obj['id']).values
8. 划分训练数据和测试数据
# 划分训练集(80%)和测试集(20%)
# CNN输入
x_train1, x_test1, y_train, y_test = train_test_split(data, lable, test_size=0.2)
# LSTM输入,标签一样,因此不用再次划分
x_train2, x_test2 = train_test_split(data, test_size=0.2)
二. 构建模型
# CNN输入格式(三维)
input_shape = (rows, cols, 1)
# input层
input_a = Input(shape=input_shape)
input_b = Input(shape=[‘LSTM输入样本序列长度’, ])
# CNN模型
model1 = Sequential([Convolution2D(32, (3, 3), activation='relu', padding='same', input_shape=input_shape),
SeparableConv2D(32, (3, 3), activation='relu', padding='same'),
MaxPooling2D(pool_size=(2, 2)),
Dropout(0.25),
Flatten()])
# LSTM模型
model2 = Sequential([Embedding(256, 32, input_length='LSTM输入样本序列长度'),
LSTM(32, dropout=0.1, return_sequences=True),
Flatten()])
# 并联操作
concat = concatenate([model1(input_a), model2(input_b)], axis=1, name="concat_layer")
# 全连接层
output = Dense(类别_num, activation='softmax')(concat) # 全连接层:每个神经元对应一个类别,输出值表示样本属于该类别的概率大小
model = Model(inputs=[input_a, input_b], outputs=[output])
三. 编译模型
# 编译模型
(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
四. 训练模型
# 训练模型
# batch_size:训练批次样本数
# epochs:训练次数
# verbose:是否打印日志
history = ([x_train1, x_train2], y_train, batch_size=128, epochs=20, verbose=2)
五. 评估模型
# 评估模型
score = ([x_test1, x_test2], y_test, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
六. 打印模型各层参数
# way1:直接利用summary()打印
()
# way2:利用plot_model()打印
import as plt
import as mpimg
from import plot_model
plot_model(model, to_file='', show_shapes=True)
photo = ('')
(photo)
('off') # 不显示坐标轴
()
七. 注:所用到的库
from sklearn import preprocessing
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from import Sequential
from keras import Model
from import Dense, Dropout, Flatten, Input
from import Convolution2D, MaxPooling2D, SeparableConv2D
from import LSTM, Embedding, concatenate