简单神经网络和卷积神经网络识别手写数字

柔光的暖阳◎ 2022-09-08 14:48 210阅读 0赞

**目录**

1 简单CNN

2 Improved CNN

3 卷积神经网络

4 参考博客和视频

5 相关函数阅读keras官方文档

--------------------

# 1 简单CNN #

实现的一个两层神经网络其隐藏层有15神元输出有10神经元

from tensorflow.keras.utils import to_categorical
    from tensorflow.keras import models, layers, regularizers
    from tensorflow.keras.optimizers import RMSprop
    from tensorflow.keras.datasets import mnist
    import matplotlib.pyplot as plt
    
    
    # 加载数据
    (train_images, train_labels), (test_images, test_labels) = mnist.load_data()
    
    # print(train_labels.shape, test_images.shape)
    # print(train_images[0])
    # print(train_labels[0])
    # plt.imshow(train_images[0])
    # plt.show()
    
    train_images = train_images.reshape((60000, 28*28)).astype("float")
    test_images = test_images.reshape((10000, 28*28)).astype("float")
    train_labels = to_categorical(train_labels)
    test_labels = to_categorical(test_labels)
    
    # print(train_images[0])
    # print(train_labels[0])
    
    network = models.Sequential()
    network.add(layers.Dense(units=15, activation='relu', input_shape=(28*28,),))
    network.add(layers.Dense(units=10, activation='softmax'))
    # 多或二分类用softmax输出的是概率 sigmod
    
    # 编译步骤
    network.compile(optimizer=RMSprop(lr=0.001), loss= 'categorical_crossentropy', metrics=["accuracy"])
    network.fit(train_images, train_labels, epochs=20, batch_size=128, verbose=2)
    
    # print(network.summary())
    # 测试
    y_pre = network.predict(test_images[:5])
    print(y_pre, test_labels[:5])
    test_loss, test_accuracy = network.evaluate(test_images, test_labels)
    print("test_loss:", test_loss, "test_accuracy:", test_accuracy)

发生过拟合了，在训练集上准确率高但在测试集上低

（每次运行结果可能不一样）

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16][]

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 1][]

网络结构

>     print(network.summary())

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 2][]

![20210826203958785.png][]

## 线性分类器 ##

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 3][]

## Softmax分类器 ##

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 4][]

先exp指数运算，好处将值变为正了，然后归一化

13%是猫 87%为车

## 交叉熵损失 ##

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 5][]

较低时效果好

# 2 Improved CNN #

from tensorflow.keras.utils import to_categorical
    from tensorflow.keras import models, layers, regularizers
    from tensorflow.keras.optimizers import RMSprop
    from tensorflow.keras.datasets import mnist
    import matplotlib.pyplot as plt
    
    
    # 加载数据
    (train_images, train_labels), (test_images, test_labels) = mnist.load_data()
    
    # print(train_labels.shape, test_images.shape)
    # print(train_images[0])
    # print(train_labels[0])
    # plt.imshow(train_images[0])
    # plt.show()
    
    train_images = train_images.reshape((60000, 28*28)).astype("float")
    test_images = test_images.reshape((10000, 28*28)).astype("float")
    train_labels = to_categorical(train_labels)
    test_labels = to_categorical(test_labels)
    
    # print(train_images[0])
    # print(train_labels[0])
    
    network = models.Sequential()
    network.add(layers.Dense(units=128, activation='relu', input_shape=(28*28,),
                             kernel_regularizer=regularizers.l1(0.0001)))
    network.add(layers.Dropout(0.01))
    network.add(layers.Dense(units=32, activation='relu',
                             kernel_regularizer=regularizers.l1(0.0001)))
    network.add(layers.Dropout(0.01))
    network.add(layers.Dense(units=10, activation='softmax'))
    # 多或二分类用softmax输出的是概率 sigmod
    
    # 编译步骤
    network.compile(optimizer=RMSprop(lr=0.001), loss= 'categorical_crossentropy', metrics=["accuracy"])
    network.fit(train_images, train_labels, epochs=20, batch_size=128, verbose=2)
    
    # print(network.summary())
    # 测试
    # y_pre = network.predict(test_images[:5])
    # print(y_pre, test_labels[:5])
    test_loss, test_accuracy = network.evaluate(test_images, test_labels)
    print("test_loss:", test_loss, "    test_accuracy:", test_accuracy)

同样发生过拟合了

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 6][]

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 7][]

网络结构

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 8][]

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_7_color_FFFFFF_t_70_g_se_x_16][]

# 3 卷积神经网络 #

from tensorflow.keras.utils import to_categorical
    from tensorflow.keras import models, layers
    from tensorflow.keras.optimizers import RMSprop
    from tensorflow.keras.datasets import mnist
    # 加载数据集
    (train_images, train_labels), (test_images, test_labels) = mnist.load_data()
    
    # 搭建LeNet网络
    def LeNet():
        network = models.Sequential()
        network.add(layers.Conv2D(filters=6, kernel_size=(3, 3), activation='relu', input_shape=(28, 28, 1)))
        network.add(layers.AveragePooling2D((2, 2)))
        network.add(layers.Conv2D(filters=16, kernel_size=(3, 3), activation='relu'))
        network.add(layers.AveragePooling2D((2, 2)))
        network.add(layers.Conv2D(filters=120, kernel_size=(3, 3), activation='relu'))
        network.add(layers.Flatten())
        network.add(layers.Dense(84, activation='relu'))
        network.add(layers.Dense(10, activation='softmax'))
        return network
    network = LeNet()
    network.compile(optimizer=RMSprop(lr=0.001), loss='categorical_crossentropy', metrics=['accuracy'])
    
    train_images = train_images.reshape((60000, 28, 28, 1)).astype('float') / 255
    test_images = test_images.reshape((10000, 28, 28, 1)).astype('float') / 255
    train_labels = to_categorical(train_labels)
    test_labels = to_categorical(test_labels)
    
    # 训练网络，用fit函数, epochs表示训练多少个回合， batch_size表示每次训练给多大的数据
    network.fit(train_images, train_labels, epochs=10, batch_size=128, verbose=2)
    test_loss, test_accuracy = network.evaluate(test_images, test_labels)
    print("test_loss:", test_loss, "    test_accuracy:", test_accuracy)

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 9][]

网络结构

![watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 10][]

# 4 参考博客和视频 #

[手把手完成mnist手写数字识别视频][mnist]

[csdn博客][csdn]

# 5 相关函数阅读keras官方文档 #

#  #

[快速开始序贯（Sequential）模型][Sequential]

[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16]: /images/20220829/00b5dcb344ac4e21a0d230364930a6ad.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 1]: /images/20220829/51baafa413a1443db773c1b2b8837033.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 2]: /images/20220829/2921a5981f4e4cc7ab5def1634260c87.png
[20210826203958785.png]: /images/20220829/9614e38f23b54f1bb2f29e33503eece2.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 3]: /images/20220829/34b041eed3404c34b99b7f6e9b57bccf.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 4]: /images/20220829/7333f593fbc648d6aef85cced457fa6b.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 5]: /images/20220829/f3c418f436ee4d57bb6b27ea3e096175.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 6]: /images/20220829/88a418eaa344453fa53a06f81d1a1742.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 7]: /images/20220829/b8da093bac504bdda2d0c64b8e8f28b1.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 8]: /images/20220829/0188afcee917434eb8e82e72f97eca6c.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_7_color_FFFFFF_t_70_g_se_x_16]: /images/20220829/b7ab17f6acdb48c9b79198b5032f6f66.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 9]: /images/20220829/614f08676b2e4aa2bb94023fded8feb1.png
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBA5aSp5Lq65ZCI5LiAcGVuZw_size_20_color_FFFFFF_t_70_g_se_x_16 10]: /images/20220829/673118797ccb446ab0b6795238e7677c.png
[mnist]: https://www.bilibili.com/video/BV16g4y1z7Qu
[csdn]: https://blog.csdn.net/qq_40438165/article/details/105846692
[Sequential]: https://keras-cn.readthedocs.io/en/latest/getting_started/sequential_model/