使用tensorflow2.x的API复现经典网络AlexNet

背景：AlexNet是2012年提出的，也是当年ImageNet竞赛当年的冠军，是一个很经典的网络结构，里面有用到卷积(C）、批归一化（B）、激活（A）、池化（P）、遗忘（D），可谓是集齐了5颗龙珠CBAPD，并且AlexNet相比较1998年卷积的开山之作的LeNet的创新之处是首次采用批归一化，激活函数从sigmoid改用relu，最后运用了Dropout。

AlexNet的网络结构：

1、卷积1、批归一化1、激活1、池化1

2、卷积2、批归一化2、激活2、池化2

3、卷积3、激活3

4、卷积4、激活4

5、卷积5、激活5、池化5

6、拉平

7、全连接1、遗忘1

8、全连接2、遗忘2

9、全连接3

import tensorflow as tf
from tensorflow.keras.layers import Conv2D, MaxPool2D, Flatten, Dense, Activation, BatchNormalization, Dropout
from tensorflow.keras import Model

# 获得数据集
cifar10 = tf.keras.datasets.cifar10
(x_train,y_train), (x_test, y_test) = cifar10.load_data()
x_train, x_test = x_train/255.0, x_test/255.0

class AlexNet(Model):
    # 构造函数
    def __init__(self):
        super(AlexNet, self).__init__()

        # 卷积
        self.c1 = Conv2D(
            filters=96,
            kernel_size=(3, 3),
            padding="valid"
        )

        # 批归一化
        self.b1 = BatchNormalization()

        # 激活
        self.a1 = Activation('relu')

        # 池化
        self.p1 = MaxPool2D(
            pool_size=(3,3),
            strides=2
        )

        
        # 卷积
        self.c2 = Conv2D(
            filters=256,
            kernel_size=(3, 3),
            padding='valid'
        )

        # 批归一化
        self.b2 = BatchNormalization()
        # 激活
        self.a2 = Activation('relu')

        # 池化
        self.p2 = MaxPool2D(
            pool_size=(3,3),
            strides=2
        )
        
        # 卷积
        self.c3 = Conv2D(
            filters=384,
            kernel_size=(3,3),
            strides=1,
            padding="same"
        )
        # 激活
        self.a3 = Activation('relu')
        # 卷积
        self.c4 = Conv2D(
            filters=384,
            kernel_size=(3, 3),
            strides=1,
            padding="same"
        )
        # 激活
        self.a4 = Activation('relu')
        # 卷积
        self.c5 = Conv2D(
            filters=256,
            kernel_size=(3, 3),
            strides=1,
            padding="same"
        )
        # 激活
        self.a5 = Activation('relu')

        # 池化
        self.p6 = MaxPool2D(
            pool_size=(3,3),
            strides=2
        )
        # 拉平
        self.f1 = Flatten()

        # 全连接
        self.d1 = Dense(2048, activation='relu')
        self.d1_ = Dropout(0.5)
        # 全连接
        self.d2 = Dense(2048, activation='relu')
        self.d2_ = Dropout(0.5)
        # 全连接
        self.d3 = Dense(10, activation='softmax')

        pass

    # 前向传播函数
    def call(self, x):
        # 卷积1、批归一化1、激活1、池化1
        x = self.c1(x)
        x = self.b1(x)
        x = self.a1(x)
        x = self.p1(x)
        # 卷积2、批归一化2、激活2、池化2
        x = self.c2(x)
        x = self.b2(x)
        x = self.a2(x)
        x = self.p2(x)
        # 卷积3、激活3
        x = self.c3(x)
        x = self.a3(x)
        # 卷积4、激活4
        x = self.c4(x)
        x = self.a4(x)
        # 卷积5、激活5、池化5
        x = self.c5(x)
        x = self.a5(x)
        x = self.p6(x)
        # 拉平
        x = self.f1(x)
        # 全连接1、遗忘1
        x = self.d1(x)
        x = self.d1_(x)
        # 全连接2、遗忘2
        x = self.d2(x)
        x = self.d2_(x)
        # 全连接3
        y = self.d3(x)
        return y

    pass


# 创建模型
model = AlexNet()

# 优化模型
model.compile(
    optimizer='adam',
    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False),
    metrics=['accuracy']
)

# 模型训练
model.fit(
    x_train,
    y_train,
    batch_size=32,
    epochs=2
)