一、理论基础

卷积神经网络

二、目的与要求

1. 掌握Keras构建卷积神经网络的主要步骤。
2. 理解卷积操作。

三、操作内容

修改如下代码，对cifar10数据库，调整网络结构为LeNet，优化算法及其学习率，批量大小batch_size，迭代的代数epoch，分析相应的结果。

要修改的代码：

# In[1]:读取数据
from keras.datasets import mnist
from keras import utils
(x_train, y_train), (x_test, y_test) = mnist.load_data()
y_train=utils.to_categorical(y_train,num_classes=10) 
y_test=utils.to_categorical(y_test,num_classes=10) 
x_train,x_test = x_train/255.0, x_test/255.0

# In[2]:构造网络
from keras import Sequential,layers,optimizers
model = Sequential( [
  layers.Reshape((28,28,1),input_shape=(28,28)),  # 二维卷积操作的输入数据要求：[样本数,宽度,高度,通道数]
  layers.Conv2D(32, kernel_size=(3, 3), activation="relu"), # 3x3的卷积核，输出32个通道
  layers.MaxPooling2D(pool_size=(2, 2)),                    # 取2x2网格的最大值进行下采样
  layers.Conv2D(64, kernel_size=(3, 3), activation="relu"),  
  layers.MaxPooling2D(pool_size=(2, 2)),
  layers.Conv2D(64, kernel_size=(3, 3), activation="relu"),
  layers.Flatten(),      # 把上一层得到的结果展平成一维向量(3*3*64=576)
  layers.Dropout(0.5),   # 训练时，每个batch随机选50%的权重固定不更新
  layers.Dense(64, activation="relu"),   
  layers.Dense(10, activation="softmax"),
])
model.summary()
#optimizer = optimizers.SGD(lr=0.5)  # 对这个例子，SGD的识别率低一点
optimizer = optimizers.Adam(lr=0.001)
#optimizer = optimizers.RMSprop(lr=0.001)
model.compile(optimizer,loss='categorical_crossentropy', metrics=['accuracy'])

# In[3]:训练和测试
model.fit(x_train, y_train, batch_size=64, epochs=1)
loss, accuracy = model.evaluate(x_test, y_test)

四、过程描述

调整网络结构为LeNet，把学习率改为0.005，批量大小改为32，迭代次数改为15次，可得到如下结果：

最后的识别率能达到0.7328。

五、源码

# -*- coding: utf-8 -*-
"""
Spyder Editor

This is a temporary script file.
"""

# In[1]:读取数据
from tensorflow.keras.datasets import cifar10
from tensorflow.keras import utils
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
y_train=utils.to_categorical(y_train,num_classes=10) 
y_test=utils.to_categorical(y_test,num_classes=10) 
x_train,x_test = x_train/255.0, x_test/255.0

# In[2]:构造网络
from tensorflow.keras import Sequential,layers,optimizers
model = Sequential( [
  layers.Reshape((32,32,3),input_shape=(32,32,3)),  # 二维卷积操作的输入数据要求：[样本数,宽度,高度,通道数]
  layers.Conv2D(32, kernel_size=(3, 3), activation="relu"), # 3x3的卷积核，输出32个通道
  layers.MaxPooling2D(pool_size=(2, 2)),                    # 取2x2网格的最大值进行下采样
  layers.Conv2D(64, kernel_size=(3, 3), activation="relu"),  
  layers.MaxPooling2D(pool_size=(2, 2)),
  layers.Conv2D(64, kernel_size=(3, 3), activation="relu"),
  layers.Flatten(),      # 把上一层得到的结果展平成一维向量(3*3*64=576)
  layers.Dropout(0.5),   # 训练时，每个batch随机选50%的权重固定不更新
  layers.Dense(64, activation="relu"),   
  layers.Dense(10, activation="softmax"),
])
model.summary()
#optimizer = optimizers.SGD(lr=0.5)  # 对这个例子，SGD的识别率低一点
optimizer = optimizers.Adam(lr=0.005)
#optimizer = optimizers.RMSprop(lr=0.001)
model.compile(optimizer,loss='categorical_crossentropy', metrics=['accuracy'])
#model.compile(optimizer,loss='mean_squared_logarithmic_error', metrics=['accuracy'])
#model.compile(optimizer,loss='mean_squared_error', metrics=['accuracy'])


# In[3]:训练和测试
#批量大小batch_size 可改为2的幂次方
model.fit(x_train, y_train, batch_size=32, epochs=15)
loss, accuracy = model.evaluate(x_test, y_test)

六、学习产出

可以通过调整了batch_size和迭代次数，就能训练出较好的结果；