经典卷积网络---LeNet、AlexNet、VGGNet、InceptionNet、ResNet [北京大学曹健老师人工智能学习笔记]

LeNet—卷积神经网络的开篇之作

Yann LeCun于1998年提出，通过共享卷积核减少了网络的参数。
LeNet有2个卷积层和3个全连接层
在特征提取阶段，卷积层以外的标准化、池化、激活等都看做是卷积层的附属，不计入层数统计。
卷积层： 6个5×5的卷积核，步长1，不使用全零填充
池化层： 2×2最大值池化，步长2，不使用全零填充

代码中只有Model类继承这里不同，完整代码参考

class LeNet(Model):def __init__(self):super(LeNet,self).__init__()  self.c1=tf.keras.layers.Conv2D(filters=6,kernel_size=(5,5),padding="valid",input_shape=(32,32,3))#该时代还未有批标准化self.a1=tf.keras.layers.Activation("sigmoid")self.p1=tf.keras.layers.MaxPool2D(pool_size=(2,2),strides=2,padding="valid")#未有dropoutself.c2=tf.keras.layers.Conv2D(filters=6,kernel_size=(5,5),padding="valid",input_shape=(32,32,3))#该时代还未有批标准化self.a2=tf.keras.layers.Activation("sigmoid")self.p2=tf.keras.layers.MaxPool2D(pool_size=(2,2),strides=2,padding="valid")self.flatten=tf.keras.layers.Flatten()self.f1=tf.keras.layers.Dense(120,activation="sigmoid")self.f2=tf.keras.layers.Dense(84,activation="sigmoid")self.f3=tf.keras.layers.Dense(10,activation="softmax")def call(self, x):#重写call,实现自己的前向传播x = self.c1(x)x = self.a1(x)x = self.p1(x)x = self.c2(x)x = self.a2(x)x = self.p2(x)x = self.flatten(x)x = self.f1(x)x = self.f2(x)y = self.f3(x)return y

AlexNet

诞生于2012年，是Hinton的代表作之一，使用relu激活函数，提高了训练速度，使用dropout[分类识别阶段]缓解了过拟合

两层3×3卷积和一层5×5卷积输出的特征图大小相同，数据比较多时，2层3×3的计算更少

AlexNet有5个卷积层和3个全连接层，共8层

class AlexNet(Model):def __init__(self):super(AlexNet,self).__init__()  self.c1=tf.keras.layers.Conv2D(filters=96,kernel_size=(3,3),padding="valid",input_shape=(32,32,3))self.b1=tf.keras.layers.BatchNormalization()   #原文为LRN，这里BN代替（更主流）self.a1=tf.keras.layers.Activation("relu")self.p1=tf.keras.layers.MaxPool2D(pool_size=(3,3),strides=2)#未有dropoutself.c2=tf.keras.layers.Conv2D(filters=256,kernel_size=(3,3),padding="valid")self.b2=tf.keras.layers.BatchNormalization()   self.a2=tf.keras.layers.Activation("relu")self.p2=tf.keras.layers.MaxPool2D(pool_size=(3,3),strides=2)self.c3=tf.keras.layers.Conv2D(filters=384,kernel_size=(3,3),padding="same")self.b3=tf.keras.layers.BatchNormalization()   self.a3=tf.keras.layers.Activation("relu")self.c4=tf.keras.layers.Conv2D(filters=384,kernel_size=(3,3),padding="same")self.b4=tf.keras.layers.BatchNormalization()   self.a4=tf.keras.layers.Activation("relu")self.c5=tf.keras.layers.Conv2D(filters=256,kernel_size=(3,3),padding="same")self.b5=tf.keras.layers.BatchNormalization()   self.a5=tf.keras.layers.Activation("relu")self.p5=tf.keras.layers.MaxPool2D(pool_size=(3,3),strides=2)self.flatten=tf.keras.layers.Flatten()self.f1=tf.keras.layers.Dense(2048,activation="relu")self.d1=tf.keras.layers.Dropout(0.5)self.f2=tf.keras.layers.Dense(2048,activation="relu")self.d2=tf.keras.layers.Dropout(0.5)self.f3=tf.keras.layers.Dense(10,activation="softmax")def call(self, x):#重写call,实现自己的前向传播x = self.c1(x)x = self.b1(x)x = self.a1(x)x = self.p1(x)x = self.c2(x)x = self.b2(x)x = self.a2(x)x = self.p2(x)x = self.c3(x)x = self.b3(x)x = self.a3(x)x = self.c4(x)x = self.b4(x)x = self.a4(x)x = self.c5(x)x = self.b5(x)x = self.a5(x)x = self.p5(x)x = self.flatten(x)x = self.f1(x)x = self.d1(x)x = self.f2(x)x = self.d2(x)y = self.f3(x)return y

VGGNet

诞生于2014年，使用小尺寸卷积核，网络结构规整，适合硬件加速

这里层数越高，卷积核的数目越多。越靠后特征图尺寸越小，通过增加卷积核的个数，增加了特征图的深度，保证了特征图的信息承载能力

16层13层卷积，3层全连接

class VGGNet(Model):def __init__(self):super(VGGNet,self).__init__()  self.c1=tf.keras.layers.Conv2D(filters=64,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b1=tf.keras.layers.BatchNormalization()self.a1=tf.keras.layers.Activation("relu")self.c2=tf.keras.layers.Conv2D(filters=64,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b2=tf.keras.layers.BatchNormalization()self.a2=tf.keras.layers.Activation("relu")self.p2=tf.keras.layers.MaxPool2D(pool_size=(2,2),strides=2,padding="same")self.d2=tf.keras.layers.Dropout(0.2)self.c3=tf.keras.layers.Conv2D(filters=128,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b3=tf.keras.layers.BatchNormalization()self.a3=tf.keras.layers.Activation("relu")self.c4=tf.keras.layers.Conv2D(filters=128,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b4=tf.keras.layers.BatchNormalization()self.a4=tf.keras.layers.Activation("relu")self.p4=tf.keras.layers.MaxPool2D(pool_size=(2,2),strides=2,padding="same")self.d4=tf.keras.layers.Dropout(0.2)self.c5=tf.keras.layers.Conv2D(filters=256,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b5=tf.keras.layers.BatchNormalization()self.a5=tf.keras.layers.Activation("relu")self.c6=tf.keras.layers.Conv2D(filters=256,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b6=tf.keras.layers.BatchNormalization()self.a6=tf.keras.layers.Activation("relu")self.c7=tf.keras.layers.Conv2D(filters=256,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b7=tf.keras.layers.BatchNormalization()self.a7=tf.keras.layers.Activation("relu")self.p7=tf.keras.layers.MaxPool2D(pool_size=(2,2),strides=2,padding="same")self.d7=tf.keras.layers.Dropout(0.2)self.c8=tf.keras.layers.Conv2D(filters=512,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b8=tf.keras.layers.BatchNormalization()self.a8=tf.keras.layers.Activation("relu")self.c9=tf.keras.layers.Conv2D(filters=512,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b9=tf.keras.layers.BatchNormalization()self.a9=tf.keras.layers.Activation("relu")self.c10=tf.keras.layers.Conv2D(filters=512,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b10=tf.keras.layers.BatchNormalization()self.a10=tf.keras.layers.Activation("relu")self.p10=tf.keras.layers.MaxPool2D(pool_size=(2,2),strides=2,padding="same")self.d10=tf.keras.layers.Dropout(0.2)self.c11=tf.keras.layers.Conv2D(filters=512,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b11=tf.keras.layers.BatchNormalization()self.a11=tf.keras.layers.Activation("relu")self.c12=tf.keras.layers.Conv2D(filters=512,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b12=tf.keras.layers.BatchNormalization()self.a12=tf.keras.layers.Activation("relu")self.c13=tf.keras.layers.Conv2D(filters=512,kernel_size=(3,3),padding="same",input_shape=(32,32,3))self.b13=tf.keras.layers.BatchNormalization()self.a13=tf.keras.layers.Activation("relu")self.p13=tf.keras.layers.MaxPool2D(pool_size=(2,2),strides=2,padding="same")self.d13=tf.keras.layers.Dropout(0.2)self.flatten=tf.keras.layers.Flatten()self.f1=tf.keras.layers.Dense(512,activation="relu")self.d14=tf.keras.layers.Dropout(0.2)self.f2=tf.keras.layers.Dense(512,activation="relu")self.d15=tf.keras.layers.Dropout(0.2)self.f3=tf.keras.layers.Dense(10,activation="softmax")def call(self, x):#重写call,实现自己的前向传播x = self.c1(x)x = self.b1(x)x = self.a1(x)x = self.c2(x)x = self.b2(x)x = self.a2(x)x = self.p2(x)x = self.d2(x)x = self.c3(x)x = self.b3(x)x = self.a3(x)x = self.c4(x)x = self.b4(x)x = self.a4(x)x = self.p4(x)x = self.d4(x)x = self.c5(x)x = self.b5(x)x = self.a5(x)x = self.c6(x)x = self.b6(x)x = self.a6(x)x = self.c7(x)x = self.b7(x)x = self.a7(x)x = self.p7(x)x = self.d7(x)x = self.c8(x)x = self.b8(x)x = self.a8(x)x = self.c9(x)x = self.b9(x)x = self.a9(x)x = self.c10(x)x = self.b10(x)x = self.a10(x)x = self.p10(x)x = self.d10(x)x = self.c11(x)x = self.b11(x)x = self.a11(x)x = self.c12(x)x = self.b12(x)x = self.a12(x)x = self.c13(x)x = self.b13(x)x = self.a13(x)x = self.p13(x)x = self.d13(x)x = self.flatten(x)x = self.f1(x)x = self.d14(x)x = self.f2(x)x = self.d14(x)y = self.f3(x)return y

InceptionNet

诞生于2014年，引入了InceptionNet结构块，同义层网络内使用不同尺寸的卷积核，提升了模型感知力，使用批标准化，缓解了梯度消失

核心：InceptionNet结构块
通过1×1卷积核，作用于输入特征图的每个像素作用点。通过设定少于输入特征图深度的1×1卷积核个数，减少输出特征图深度，起到降维的作用，减少参数量和计算量

包含四个分支【没备注的都是卷积运算】
1×1
1×1 3×3
1×1 5×5
3×3(池化) 1×1
四个分支输出的特征数据尺寸相同，送入卷积连接器，卷积连接器将收到的数据按深度方向拼接，形成InceptionNet结构块输出

第一层为16个3*3卷积核，4个InceptionNet结构块,2个组成一个block,平均值池化，分类

class ConBNRelu(Model):def __init__(self,ch,kernel_size=3,strides=1,padding="same"):super(ConBNRelu,self).__init__()self.model=tf.keras.models.Sequential([tf.keras.layers.Conv2D(ch,kernel_size,strides=strides,padding=padding),tf.keras.layers.BatchNormalization(),tf.keras.layers.Activation("relu")])def call(self,x):x=self.model(x)return x     class InceptionBLK(Model):def __init__(self,ch,strides=1):super(InceptionBLK,self).__init__()self.ch=chself.strides=stridesself.c1=ConBNRelu(ch,kernel_size=1,strides=strides)self.c2_1=ConBNRelu(ch,kernel_size=1,strides=strides)self.c2_2=ConBNRelu(ch,kernel_size=3,strides=1)self.c3_1=ConBNRelu(ch,kernel_size=1,strides=strides)self.c3_2=ConBNRelu(ch,kernel_size=5,strides=1)self.p4_1=tf.keras.layers.MaxPool2D(3,strides=1,padding="same")self.c4_2=ConBNRelu(ch,kernel_size=1,strides=strides)def call(self,x):x1=self.c1(x)x2_1=self.c2_1(x)x2_2=self.c2_2(x2_1)x3_1=self.c3_1(x)x3_2=self.c3_2(x3_1)x4_1=self.p4_1(x)x4_2=self.c4_2(x4_1)x=tf.concat([x1,x2_2,x3_2,x4_2],axis=3)return x     class InceptionNet(Model):def __init__(self,num_block,num_classes,init_ch=16,**kwargs):super(InceptionNet,self).__init__(**kwargs)self.inchannel=init_chself.outchannel=init_chself.init_ch=init_chself.c1=ConBNRelu(init_ch)self.blocks=tf.keras.models.Sequential()for block_id in range(num_block):for layer_id in range(2):if layer_id==0:block=InceptionBLK(self.outchannel,strides=2) #第一个InceptionNet结构块步长为2，使输出特征图尺寸减半else:block=InceptionBLK(self.outchannel,strides=1)self.blocks.add(block)self.outchannel*=2self.p1=tf.keras.layers.GlobalAveragePooling2D()self.f1=tf.keras.layers.Dense(num_classes,activation="softmax")def call(self,x):x=self.c1(x)x=self.blocks(x)x=self.p1(x)y=self.f1(x)return ymodel=InceptionNet(num_block=2,num_classes=10)

ResNet

诞生于2015年，提出了层间残差跳连，引入前方信息，缓解梯度消失，使神经网络层数增加成为可能。resnet作者何凯明发现单纯堆叠神经网络层数，会使神经网络模型退化，以至于后面的特征丢失了前边特征的原本模样。
他用跳连线，将前面的特征直接接到了后面。ResNet块中的+是特征图对应元素值相加。

ResNet块有两种形式
堆叠卷积前后维度相同，可以直接相加
堆叠卷积前后维度不同，通过1×1卷积核改变维度

这里我跟着老师写的代码有问题，训练结果准确率一直是0.1左右，就从别的地方copy了一下

import tensorflow as tf
physical_devices = tf.config.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0], True)
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.keras import Model
from tensorflow.keras.layers import Conv2D,BatchNormalization,Activation,GlobalAveragePooling2D,Dense
import pandas as pd(x_train,y_train),(x_test,y_test)=tf.keras.datasets.cifar10.load_data()x_train,x_test=x_train/255.0,x_test/255.0class ResnetBlock(Model):def __init__(self, filters, strides=1, residual_path=False):super(ResnetBlock, self).__init__()self.filters = filtersself.strides = stridesself.residual_path = residual_pathself.c1 = Conv2D(filters, (3, 3), strides=strides, padding='same', use_bias=False)self.b1 = BatchNormalization()self.a1 = Activation('relu')self.c2 = Conv2D(filters, (3, 3), strides=1, padding='same', use_bias=False)self.b2 = BatchNormalization()# residual_path为True时，对输入进行下采样，即用1x1的卷积核做卷积操作，保证x能和F(x)维度相同，顺利相加if residual_path:self.down_c1 = Conv2D(filters, (1, 1), strides=strides, padding='same', use_bias=False)self.down_b1 = BatchNormalization()self.a2 = Activation('relu')def call(self, inputs):residual = inputs  # residual等于输入值本身，即residual=x# 将输入通过卷积、BN层、激活层，计算F(x)x = self.c1(inputs)x = self.b1(x)x = self.a1(x)x = self.c2(x)y = self.b2(x)if self.residual_path:residual = self.down_c1(inputs)residual = self.down_b1(residual)out = self.a2(y + residual)  # 最后输出的是两部分的和，即F(x)+x或F(x)+Wx,再过激活函数return outclass ResNet18(Model):def __init__(self, block_list, initial_filters=64):  # block_list表示每个block有几个卷积层super(ResNet18, self).__init__()self.num_blocks = len(block_list)  # 共有几个blockself.block_list = block_listself.out_filters = initial_filtersself.c1 = Conv2D(self.out_filters, (3, 3), strides=1, padding='same', use_bias=False)self.b1 = BatchNormalization()self.a1 = Activation('relu')self.blocks = tf.keras.models.Sequential()# 构建ResNet网络结构for block_id in range(len(block_list)):  # 第几个resnet blockfor layer_id in range(block_list[block_id]):  # 第几个卷积层if block_id != 0 and layer_id == 0:  # 对除第一个block以外的每个block的输入进行下采样block = ResnetBlock(self.out_filters, strides=2, residual_path=True)else:block = ResnetBlock(self.out_filters, residual_path=False)self.blocks.add(block)  # 将构建好的block加入resnetself.out_filters *= 2  # 下一个block的卷积核数是上一个block的2倍self.p1 = tf.keras.layers.GlobalAveragePooling2D()self.f1 = tf.keras.layers.Dense(10, activation='softmax', kernel_regularizer=tf.keras.regularizers.l2())def call(self, inputs):x = self.c1(inputs)x = self.b1(x)x = self.a1(x)x = self.blocks(x)x = self.p1(x)y = self.f1(x)return y       model=ResNet18([2,2,2,2])model.compile(optimizer="adam",loss="sparse_categorical_crossentropy",metrics="sparse_categorical_accuracy")checkpoint_save_path = "./VGGNet_cifar10.ckpt"
cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_save_path, save_weights_only=True, save_best_only=True)history=model.fit(x_train,y_train,batch_size=400,epochs=5,validation_data=(x_test,y_test),validation_freq=1,callbacks=[cp_callback])
pd.DataFrame(history.history).to_csv("cifar_training_log.csv",index=False)graph=pd.read_csv("cifar_training_log.csv")
graph.plot(figsize=(5,4))
plt.xlim(0,4)
plt.ylim(0,2)
plt.grid(1)plt.show()num=np.random.randint(1,10000)
demo=tf.reshape(x_test[num],(1,32,32,3))
y_pred=np.argmax(model.predict(demo))
plt.imshow(x_test[num])
plt.show()
print("标签值："+str(y_test[num,0])+"\n预测值:"+str(y_pred))
#飞机，汽车，鸟，猫，鹿，狗，青蛙，马，船，卡车

运行结果比较

LeNet

AlexNet

VGGNet

InceptionNet

ResNet