用Tensorflow2搭建经典的卷积神经网络（LeNet,AlexNet,VGG16,ResNet50）

最新推荐文章于 2025-06-12 15:28:56 发布

原创

最新推荐文章于 2025-06-12 15:28:56 发布 · 1.5k 阅读

16 ·

CC 4.0 BY-SA版权

文章标签：

#神经网络 #tensorflow #图像识别 #深度学习

本文介绍了使用Tensorflow2复现经典卷积神经网络的过程，包括LeNet、AlexNet、VGG16和ResNet50，并利用mnist数据集测试LeNet。实验结果显示网络搭建正确，ResNet50在层数最多的情况下训练参数最少。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

最近看了卷积神经网络的一些经典论文，用Tensorflow2进行了复现，其中包括LeNet,AlexNet,VGG16,ResNet50,并且在数据集上进行了测试，能够正常训练，结果表明搭建的没有问题。

下面进入正题：

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline

加载mnist数据集用于测试LeNet

mnist_data = np.load("E:/deep_learning/jupyternotebook工作空间/0531/mnist/mnist.npz")
x_train = mnist_data['x_train'] / 255.0
y_train = mnist_data['y_train'] / 255.0
x_test = mnist_data['x_test']
y_test = mnist_data['y_test']

print(x_train.shape)
print(x_test.shape)
print(y_train.shape)
print(y_test.shape)

(60000, 28, 28)
(10000, 28, 28)
(60000,)
(10000,)

x_train0 = x_train.reshape((-1,28,28,1))
x_test0 = x_test.reshape((-1,28,28,1))

plt.imshow(x_test[0])
plt.title('The number is '+ str(y_test[0]))

Text(0.5, 1.0, 'The number is 7')

在这里插入图片描述

LeNet

LeNet模型网络结构示意图

def get_LeNet():
    model = keras.Sequential(name='LeNet')
    model.add(layers.Conv2D(filters=6,kernel_size=(5,5),strides=1,activation='sigmoid',
                            input_shape=(28,28,1),name='Conv1'))
    model.add(layers.MaxPool2D(pool_size=(2,2),strides=2,name='Pool1'))
    
    model.add(layers.Conv2D(16,(5,5),strides=1,activation='sigmoid',name='Conv2'))
    model.add(layers.MaxPool2D(pool_size=(2,2),strides=2,name='Pool2'))
    
    model.add(layers.Conv2D(120,(4,4),strides=1,activation='sigmoid',name='Conv3'))
    
    model.add(layers.Flatten(name='Flatten'))
    model.add(layers.Dense(64,activation='sigmoid',name='FC1'))
    model.add(layers.Dense(10,activation='softmax',name='FC2'))
    return model

LeNet = get_LeNet()
LeNet.summary()

Model: "LeNet"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
Conv1 (Conv2D)               (None, 24, 24, 6)         156       
_________________________________________________________________
Pool1 (MaxPooling2D)         (None, 12, 12, 6)         0         
_________________________________________________________________
Conv2 (Conv2D)               (None, 8, 8, 16)          2416      
_________________________________________________________________
Pool2 (MaxPooling2D)         (None, 4, 4, 16)          0         
_________________________________________________________________
Conv3 (Conv2D)               (None, 1, 1, 120)         30840     
_________________________________________________________________
Flatten (Flatten)            (None, 120)               0         
_________________________________________________________________
FC1 (Dense)                  (None, 64)                7744      
_________________________________________________________________
FC2 (Dense)                  (None, 10)                650       
=================================================================
Total params: 41,806
Trainable params: 41,806
Non-trainable params: 0
_________________________________________________________________

由上图可知，LeNet为5层网络，共41806个参数

#训练
LeNet.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001),
                   loss='sparse_categorical_crossentropy',   #计算稀疏的分类交叉熵损失
                    metrics=['accuracy'])

LeNet.fit(x_train0,y_train,batch_size=64,epochs=1,validation_split=0.15)
LeNet.evaluate(x_test0,y_test,batch_size=64)

AlexNet

在这里插入图片描述

def alexNet_model(numclasses):
    alexNet = keras.Sequential(name='AlexNet')
    #第一层
    alexNet.add(layers.Conv2D(filters=96,kernel_size=(11,11),strides=(4,4),
                              activation='relu',input_shape=(227,227,3),name='Conv1'))
    alexNet.add(layers.BatchNormalization(name='BN1'))
    alexNet.add(layers.MaxPool2D(pool_size=(3,3),strides=(2,2),name='Pool1'))
    
    #第二层
    alexNet.add(layers.Conv2D(filters=256,kernel_size=(5,5),strides=(1,1),padding='same',
                              activation='relu',name='Conv2'))
    alexNet.add(layers.BatchNormalization(name='BN2'))
    alexNet.add(layers.MaxPool2D(pool_size=(3,3),strides=(2,2),name='Pool2'))
    
    #第三层
    alexNet.add(layers.Conv2D(filters=384,kernel_size=(3,3),strides=(1,1),padding='same',
                              activation='relu',name='Conv3'))
    
    #第四层
    alexNet.add(layers.Conv2D(filters=192,kernel_size=(3,3),strides=(1,1),padding='same',
                              activation='relu',name='Conv4'))
    
    #第五层
    alexNet.add(layers.Conv2D(filters=256,kernel_size=(3,3),strides=(1,1),padding=