Keras搭建经典CNN LeNet5网络进行手写体识别

1 简介

LeNet-5 模型是Yann LeCun 教授于 1998 年在论文 Gradient-based learning applied to document recognition中提出的，它是第一个成功应用于数字识别问题的卷积神经网络。在 MNIST 数据集上， LeNet-5 模型可以达到大约 **99.2%**的正确率。

LeNet-5的网络结构如下：

在minst手写体数据集基础上构建LeNet-5网络，其结构如下：

其中，T表示tanh激活函数，S代表softmax。

2 数据处理和展示

原始mnist数据格式为**[6000,28,28]，但是卷积层所需要的格式为[None,28,28,1]，None 代表任何个[28,28,1]数据，其中28为宽高**，1代表颜色通道（一般颜色通道为3，rgb，这里采用灰色).

 def load_data(self):  
        print('loaddata')  
        (X_train, y_train), (X_test, y_test) = mnist.load_data()  
        X_train = X_train.reshape(-1, 28, 28, 1)  # normalize  
 		X_test = X_test.reshape(-1, 28, 28, 1)  # normalize  
		X_train = X_train / 255  
 		X_test = X_test / 255  
 		y_train = to_categorical(y_train, num_classes=10)  
        y_test = to_categorical(y_test, num_classes=10)  
        return (X_train, y_train), (X_test, y_test)  

我们选取了数据中9张图进行展示：

3 模型搭建

其中LeNet5中的5代表5层，我已经在下面用彩色数字标注出来了。

4 代码实现

import tensorflow as tf  
from tensorflow.keras.datasets import mnist  
from tensorflow.keras.layers import Dense, Conv2D, AveragePooling2D, Flatten  
from tensorflow.keras.models import Sequential  
from tensorflow.keras.utils import to_categorical  

class LeNet5:  
    def __init__(self):  
        self.model = Sequential()  
  
    def load_data(self):  
        print('loaddata')  
        (X_train, y_train), (X_test, y_test) = mnist.load_data()  
        X_train = X_train.reshape(-1, 28, 28, 1)  # normalize  
 		X_test = X_test.reshape(-1, 28, 28, 1)  # normalize  
		X_train = X_train / 255  
 		X_test = X_test / 255  
 		y_train = to_categorical(y_train, num_classes=10)  
        y_test = to_categorical(y_test, num_classes=10)  
        return (X_train, y_train), (X_test, y_test)  
  
    def train(self, x, y):  
        print("build")  
        self.model.add(Conv2D(input_shape=(28, 28, 1),  
                              filters=6,  
                              kernel_size=(5, 5),  
                              strides=1,  
                              activation=tf.keras.activations.tanh))  
        self.model.add(AveragePooling2D(pool_size=(2, 2)))  
        self.model.add(Conv2D(filters=6,  
                              kernel_size=(5, 5),  
                              strides=1,  
                              activation=tf.keras.activations.tanh))  
        self.model.add(AveragePooling2D(pool_size=(2, 2)))  
        self.model.add(Flatten())  
        self.model.add(Dense(120, activation=tf.keras.activations.tanh))  
        self.model.add(Dense(84, activation=tf.keras.activations.tanh))  
        self.model.add(Dense(10, activation=tf.keras.activations.softmax)) 
		
        # Compile model  
 		self.model.compile(loss=tf.keras.losses.CategoricalCrossentropy(),  
                           optimizer="adam",  
                           metrics=['accuracy'])  
        print("train")  
        self.model.fit(x=x, y=y, batch_size=1, epochs=100, verbose=1)  
        return self.model  
  
    def predict(self, x):  
        print("predict")  
        for result in self.model.predict(x):  
            print(tf.argmax(result))  
  
  
if __name__ == "__main__":  
    lenet = LeNet5()  
    (X_train, y_train), (X_test, y_test) = lenet.load_data()  
  
    lenet.train(X_train, y_train)  
    lenet.predict(X_test)

5 评估

模型结构：

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 24, 24, 6)         156       
_________________________________________________________________
average_pooling2d (AveragePo (None, 12, 12, 6)         0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 8, 8, 6)           906       
_________________________________________________________________
average_pooling2d_1 (Average (None, 4, 4, 6)           0         
_________________________________________________________________
flatten (Flatten)            (None, 96)                0         
_________________________________________________________________
dense (Dense)                (None, 120)               11640     
_________________________________________________________________
dense_1 (Dense)              (None, 84)                10164     
_________________________________________________________________
dense_2 (Dense)              (None, 10)                850       
=================================================================
Total params: 23,716
Trainable params: 23,716
Non-trainable params: 0
_________________________________________________________________

为了快速得到结果，设置epoch=1得到训练结果：

loss: 0.1207 - accuracy: 0.9625