0%

TensorFlow实践之典型的卷积神经网路

发表于 2022-11-07 更新于 2022-11-11 分类于学习积累阅读次数：

本节主要介绍了LeNet、VGGNet、GoogLeNet神经网络。其中，对LeNet模型进行了函数封装，对VGGNet进行了基本介绍，并着重介绍了GoogLeNet神经网络中1×1卷积核的作用。

LeNet神经网络

1.网络结构

2.封装LeNet模型

def LeNet(input_shape,padding):
    model=tf.keras.Sequential()
    #添加卷积层1、池化层1
    model.add(tf.keras.layers.Conv2D(6,kernel_size=(5,5),padding=padding,activation="sigmoid",input_shape=input_shape))
    model.add(tf.keras.layers.MaxPool2D(pool_size=(2,2)))
    #添加卷积层2、池化层2
    model.add(tf.keras.layers.Conv2D(16,kernel_size=(5,5),padding="valid",activation="sigmoid"))
    model.add(tf.keras.layers.MaxPool2D(pool_size=(2,2)))
    #添加卷积层3
    model.add(tf.keras.layers.Conv2D(120,kernel_size=(5,5),padding="valid",activation="sigmoid"))#本层输出为形状为（1，1，120）
    #添加Flatten层，将数据拉直
    model.add(tf.keras.layers.Flatten())
    #添加隐含层和输出层
    model.add(tf.keras.layers.Dense(84,activation="sigmoid"))
    model.add(tf.keras.layers.Dense(10,activation="softmax"))
    return model
#调用方法
'''model=LeNet((28,28,1),"same")
model=LeNet((32,32,1),"valid")'''

3.实例：LeNet实现手写数字识别

#1.导入库
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
#2.加载数据
mnist=tf.keras.datasets.mnist
(train_x,train_y),(test_x,test_y)=mnist.load_data(path="D:\App_Data_File\Anaconda_data\jupyter\image\mnist.npz")
#3.数据预处理:对属性进行归一化并转化为Tensor张量
X_train,X_test=tf.cast(train_x/255.0,tf.float32),tf.cast(test_x/255.0,tf.float32)
Y_train,Y_test=tf.cast(train_y,tf.int16),tf.cast(test_y,tf.int16)
X_train=tf.reshape(X_train,[60000,28,28,1])
X_test=tf.reshape(X_test,[10000,28,28,1])
#4.建立LeNet模型
'''model=tf.keras.Sequential()
#添加卷积层1、池化层1
model.add(tf.keras.layers.Conv2D(6,kernel_size=(5,5),padding="same",activation="sigmoid",input_shape=(28,28,1)))
model.add(tf.keras.layers.MaxPool2D(pool_size=(2,2)))
#添加卷积层2、池化层2
model.add(tf.keras.layers.Conv2D(16,kernel_size=(5,5),padding="valid",activation="sigmoid"))
model.add(tf.keras.layers.MaxPool2D(pool_size=(2,2)))
#添加卷积层3
model.add(tf.keras.layers.Conv2D(120,kernel_size=(5,5),padding="valid",activation="sigmoid"))
#添加Flatten层，将数据拉直
model.add(tf.keras.layers.Flatten())
#添加隐含层和输出层
model.add(tf.keras.layers.Dense(84,activation="sigmoid"))
model.add(tf.keras.layers.Dense(10,activation="softmax"))'''
model=LeNet((28,28,1),"same")
model.summary()
#5.配置训练方法
model.compile(optimizer="adam",loss="sparse_categorical_crossentropy",metrics=["accuracy"])
#6.训练模型
history=model.fit(X_train,Y_train,batch_size=64,epochs=5,validation_split=0.2)
#7.评估模型
model.evaluate(X_test,Y_test,verbose=2)
#8.使用模型
# 下面再随机取出测试集中的任意10个数据进行识别
plt.rcParams["font.sans-serif"]="SimHei"
plt.rcParams["axes.unicode_minus"]=False
plt.figure()
plt.suptitle("随机取出测试集中的任意10个数据进行识别", fontsize=20, color="blue")
for i in range(10):
    num = np.random.randint(1, 10000)
    plt.subplot(2, 5, i+1)
    plt.axis("off")
    plt.imshow(test_x[num], cmap="gray")
    y_pred = np.argmax(model.predict(tf.reshape(X_test[num],(1,28,28,1))))
    plt.title("y= "+str(test_y[num]) + "\n" + "y_pred=" + str(y_pred))
plt.tight_layout(rect=[0,0,1,0.98])
plt.show()

结果如下：

Model: "sequential_9"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_27 (Conv2D)           (None, 28, 28, 6)         156       
_________________________________________________________________
max_pooling2d_18 (MaxPooling (None, 14, 14, 6)         0         
_________________________________________________________________
conv2d_28 (Conv2D)           (None, 10, 10, 16)        2416      
_________________________________________________________________
max_pooling2d_19 (MaxPooling (None, 5, 5, 16)          0         
_________________________________________________________________
conv2d_29 (Conv2D)           (None, 1, 1, 120)         48120     
_________________________________________________________________
flatten_9 (Flatten)          (None, 120)               0         
_________________________________________________________________
dense_18 (Dense)             (None, 84)                10164     
_________________________________________________________________
dense_19 (Dense)             (None, 10)                850       
=================================================================
Total params: 61,706
Trainable params: 61,706
Non-trainable params: 0
_________________________________________________________________
Epoch 1/5
750/750 [==============================] - 23s 29ms/step - loss: 1.1872 - accuracy: 0.6143 - val_loss: 0.3047 - val_accuracy: 0.9153
Epoch 2/5
750/750 [==============================] - 22s 29ms/step - loss: 0.2160 - accuracy: 0.9393 - val_loss: 0.1392 - val_accuracy: 0.9581
Epoch 3/5
750/750 [==============================] - 22s 29ms/step - loss: 0.1238 - accuracy: 0.9627 - val_loss: 0.0956 - val_accuracy: 0.9721
Epoch 4/5
750/750 [==============================] - 21s 28ms/step - loss: 0.0919 - accuracy: 0.9717 - val_loss: 0.0836 - val_accuracy: 0.9755
Epoch 5/5
750/750 [==============================] - 21s 28ms/step - loss: 0.0759 - accuracy: 0.9766 - val_loss: 0.0788 - val_accuracy: 0.9761
313/313 - 3s - loss: 0.0719 - accuracy: 0.9774

1 2	#保存训练的模型参数 model.save_weights("mnist_save_weights_LeNet.h5")

AlexNet神经网络

AlexNet网络结构
AlexNet创新点
- 使用ReLU激活函数：使网络训练收敛更快
- 使用Dropout训练模型
- 数据增强：增加了2048倍的数据量
  - 原始图像256×256，随机截取224×224区域
  - 水平翻转的镜像
- 使用重叠最大池化：能够很好克服过拟合问题
- 使用GPU加速训练过程

VGGNet神经网络

多个小卷积核级联加深网络结构
- 减少网络计算量
- 增加网络的泛化能力和表达能力
先训练级别A的简单网络，再复用A网络的权重初始化复杂模型

#VGG16模型的建立
model=tf.keras.Sequential()
#添加卷积层1-2、池化层
model.add(tf.keras.layers.Conv2D(64,kernel_size=(3,3),padding="same",activation=tf.nn.relu,input_shape=(224,244,3)))
model.add(tf.keras.layers.Conv2D(64,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.MaxPool2D(pool_size=(2,2)))
#添加卷积层3-4、池化层
model.add(tf.keras.layers.Conv2D(128,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.Conv2D(128,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.MaxPool2D(pool_size=(2,2)))
#添加卷积层5-7、池化层
model.add(tf.keras.layers.Conv2D(256,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.Conv2D(256,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.Conv2D(256,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.MaxPool2D(pool_size=(2,2)))
#添加卷积层8-10、池化层
model.add(tf.keras.layers.Conv2D(512,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.Conv2D(512,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.Conv2D(512,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.MaxPool2D(pool_size=(2,2)))
#添加卷积层11-13、池化层
model.add(tf.keras.layers.Conv2D(512,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.Conv2D(512,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.Conv2D(512,kernel_size=(3,3),padding="same",activation=tf.nn.relu))
model.add(tf.keras.layers.MaxPool2D(pool_size=(2,2)))
#添加Flatten层，将数据拉直
model.add(tf.keras.layers.Flatten())
#全连接层1，Dropout层
model.add(tf.keras.layers.Dense(4096,activation="relu"))
model.add(tf.keras.layers.Dropout(0.5))
#全连接层2，Dropout层
model.add(tf.keras.layers.Dense(4096,activation="relu"))
model.add(tf.keras.layers.Dropout(0.5))
#全连接层3（输出层）
model.add(tf.keras.layers.Dense(1000,activation="softmax"))
model.summary()

结果如下：

Model: "sequential_14"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_56 (Conv2D)           (None, 224, 244, 64)      1792      
_________________________________________________________________
conv2d_57 (Conv2D)           (None, 224, 244, 64)      36928     
_________________________________________________________________
max_pooling2d_32 (MaxPooling (None, 112, 122, 64)      0         
_________________________________________________________________
conv2d_58 (Conv2D)           (None, 112, 122, 128)     73856     
_________________________________________________________________
conv2d_59 (Conv2D)           (None, 112, 122, 128)     147584    
_________________________________________________________________
max_pooling2d_33 (MaxPooling (None, 56, 61, 128)       0         
_________________________________________________________________
conv2d_60 (Conv2D)           (None, 56, 61, 256)       295168    
_________________________________________________________________
conv2d_61 (Conv2D)           (None, 56, 61, 256)       590080    
_________________________________________________________________
conv2d_62 (Conv2D)           (None, 56, 61, 256)       590080    
_________________________________________________________________
max_pooling2d_34 (MaxPooling (None, 28, 30, 256)       0         
_________________________________________________________________
conv2d_63 (Conv2D)           (None, 28, 30, 512)       1180160   
_________________________________________________________________
conv2d_64 (Conv2D)           (None, 28, 30, 512)       2359808   
_________________________________________________________________
conv2d_65 (Conv2D)           (None, 28, 30, 512)       2359808   
_________________________________________________________________
max_pooling2d_35 (MaxPooling (None, 14, 15, 512)       0         
_________________________________________________________________
conv2d_66 (Conv2D)           (None, 14, 15, 512)       2359808   
_________________________________________________________________
conv2d_67 (Conv2D)           (None, 14, 15, 512)       2359808   
_________________________________________________________________
conv2d_68 (Conv2D)           (None, 14, 15, 512)       2359808   
_________________________________________________________________
max_pooling2d_36 (MaxPooling (None, 7, 7, 512)         0         
_________________________________________________________________
flatten_14 (Flatten)         (None, 25088)             0         
_________________________________________________________________
dense_30 (Dense)             (None, 4096)              102764544 
_________________________________________________________________
dropout_4 (Dropout)          (None, 4096)              0         
_________________________________________________________________
dense_31 (Dense)             (None, 4096)              16781312  
_________________________________________________________________
dropout_5 (Dropout)          (None, 4096)              0         
_________________________________________________________________
dense_32 (Dense)             (None, 1000)              4097000   
=================================================================
Total params: 138,357,544
Trainable params: 138,357,544
Non-trainable params: 0
_________________________________________________________________

GoogLeNet神经网络

1.Inception Module的初代

2.1×1卷积核

能够降维/升维，跨通道信息交流
个人对1×1，7与1×1×7，1的理解：前者表示有七个1×1的卷积核，其中每个卷积核都与图像去卷一卷，那么每个卷积核都会输出一个特征图，则一个会输出七个特征图；后者表示一个卷积核，这个卷积核的大小是1×1×7，可以理解为有七个通道，每个通道里都有1×1的一个小卷积核（也就是这7个所谓的小卷积核从不同维度共同组成一个整体的大卷积核，实际上还是只有一个卷积核），这7个小卷积核与前一层中7个对应通道的图像分别卷积，将其结果相加后，最终还是只会得到一个特征图。

欢迎来到ssy的世界

本文作者： ssy
本文链接： http://ssy的小天地.com/2022/11/07/TensorFlow实践之典型的卷积神经网络/
版权声明： 本博客所有文章除特别声明外，均采用 BY-NC-SA 许可协议。转载请注明出处！