Keras+tensorboard实现mnist手写数字识别（附完整代码）

Keras+tensorboard实现mnist手写数字识别

此次利用Keras的Model函数式编程方式完成框架的搭建，利用tensorboard来辅助生成各种变量的数据变化图。其中也会用到一些训练和图像处理的小技巧，下文会详尽描述。（文末附完整代码）

1. 导入所需包

from __future__ import print_function, division

from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout
from keras.layers import BatchNormalization, Activation, ZeroPadding2D, Conv2D, Conv2DTranspose,MaxPooling2D
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import UpSampling2D
from keras.models import Sequential, Model
from keras.optimizers import Adam
from keras.utils.np_utils import to_categorical
import matplotlib.pyplot as plt
from keras.preprocessing.image import ImageDataGenerator
from sklearn.model_selection import train_test_split
from keras.callbacks import ReduceLROnPlateau
import sys

import numpy as np

2. 载入mnist数据集并进行预处理

2.1 使用from keras.datasets import mnist载入数据集

(x_train, y_train),(x_test, y_test) = mnist.load_data()

因为我们2D卷积网络的输入需要四个维度（batch_size,hight,weigh,channels），其中：

batch_size 代表每次输入的批次大小
hight 代表图片的高度
weigh 代表图片的宽度
channels 代表信道数目

2.2 归一化处理

所以我们需要将其转化为（batch_size,hight,weigh,channels）这样的形式，并，将数据归一化到（0,1）之间。

batch_size = 64 #每次传入的批次大小
epochs = 30 #训练次数
(x_train, y_train),(x_test, y_test) = mnist.load_data() #载入数据集
x_train = np.expand_dims(x_train,axis=3) / 255. #扩充维度并进行归一化处理
x_test = np.expand_dims(x_test,axis=3) / 255.

2.3 将标签进行one-hot表示

我们的数据集包含0-9这是个数字，所以需要进行10分类，这也就确定了我们神经网络的最后的输出是(batch_size,10).

我们的标签y则需要转化成类似于[0,1,0,0,0,0,0,0,0,0]的one-hot编码的形式，这里我们使用到to_categorical的函数，此函数作用是to_categorical就是将类别向量转换为二进制（只有0和1）的矩阵类型表示。其表现为将原有的类别向量转换为独热编码的形式。

y_train = to_categorical(y_train, num_classes = 10)#将标签转换成类似[0,1,0,0,0,0,0,0,0,0] one-hot编码，其中num_classes 代表类别个数
y_test = to_categorical(y_test, num_classes = 10)

2.4 进行训练集验证机的划分（使用train_test_split）

x_train, x_val, y_train, y_val = train_test_split(x_train, y_train, test_size = 0.1, random_state=42)

3. 使用keras搭建网络

3.1 搭建神经网络

#开始构建网络结构
#初始化 Input
input_ = Input(shape=(28,28,1,))
#定义卷积层、池化层、Dropout层
x = Conv2D(filters=32, kernel_size=(3, 3),activation='relu',kernel_initializer='he_normal')(input_)
x = Conv2D(32, kernel_size=(3, 3),activation='relu',kernel_initializer='he_normal')(x)
x = MaxPooling2D((2, 2))(x)
x = Dropout(0.20)(x)

x = Conv2D(64, (3, 3), activation='relu',padding='same',kernel_initializer='he_normal')(x)
x = Conv2D(64, (3, 3), activation='relu',padding='same',kernel_initializer='he_normal')(x)
x = MaxPooling2D(pool_size=(2, 2))(x)
x = Dropout(0.25)(x)

x = Conv2D(128, (3, 3), activation='relu',padding='same',kernel_initializer='he_normal')(x)
x = Dropout(0.25)(x)
#从这里转为全连接神经网络
x = Flatten()(x)
x = Dense(128, activation='relu')(x)
x= BatchNormalization()(x)
x = Dropout(0.25)(x)
output= Dense(10,activation='softmax')(x)

model = Model(input_, output)#初始化模型
model.summary()#输出模型参数
model.compile(optimizer='Adam', #编译模型
              loss = 'categorical_crossentropy',
              metrics=['accuracy']
             )

编译结果为：

3.2 将图片进行变换

这个操作是为了增强模型的泛化能力，同时生成更多的训练样本。通过生成这样的生成器可以进行小批量的输入，从而减少内存的占用。这里我们对每个批次的张图片进行了15度的旋转、水平和竖直方向各移动了总像素点的0.1倍。
通过使用ImageDataGenerator来生成：

datagen = ImageDataGenerator(
        featurewise_center=False,  # set input mean to 0 over the dataset
        samplewise_center=False,  # set each sample mean to 0
        featurewise_std_normalization=False,  # divide inputs by std of the dataset
        samplewise_std_normalization=False,  # divide each input by its std
        zca_whitening=False,  # apply ZCA whitening
        rotation_range=15, # randomly rotate images in the range (degrees, 0 to 180)
        zoom_range = 0.1, # Randomly zoom image 
        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
        height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)
        horizontal_flip=False,  # randomly flip images
        vertical_flip=False)  # randomly flip images

datagen.fit(x_train)

3.3 利用TensorBoard来保存日志

tbCallBack = TensorBoard(log_dir="./model"，histogram_freq=5,write_grads=True)  #将日志文件放到model文件下

3.4 使用ReduceLROnPlateau来进行训练过程中的学习率更新

这个函数的作用是当在确定的训练批次内我们的监测指标不在提升，则降低我们的学习率

learning_rate_reduction = ReduceLROnPlateau(monitor='val_acc', #监测指标
                                            patience=3, #3个epochs内指标不提升则更新学习率
                                            verbose=1, #显示输出
                                            factor=0.5, #lr = 0.5*lr
                                            min_lr=0.0001)#学习率的最小值

3.5 开始训练和测试模型

history = model.fit_generator(datagen.flow(x_train,y_train, batch_size=batch_size),
                              epochs = epochs, validation_data = (x_val,y_val),
                              verbose = 1, steps_per_epoch=x_train.shape[0] // batch_size
                              , callbacks=[learning_rate_reduction,tbCallBack])

model.evaluate(x_test, y_test)

4. 最终结果

4.1 训练结果

loss: 0.0168 - acc: 0.9947 - val_loss: 0.0154 - val_acc: 0.9957

4.2 测试集结果

测试集的准确率到达0.9958

[0.012207253835405572, 0.9958]

4.3 tensorboard可视化—结构图

4.3 tensorboard可视化—loss、acc、lr

4.3.1 训练集的acc、loss：

4.3.2 lr

4.3.3 测试集的acc、loss

4.3 tensorboard可视化—权重和偏置值（以BN的为例）

5. 完整代码

from __future__ import print_function, division

from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout
from keras.layers import BatchNormalization, Activation, ZeroPadding2D, Conv2D, Conv2DTranspose,MaxPooling2D
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import UpSampling2D
from keras.models import Sequential, Model
from keras.optimizers import Adam
from keras.utils.np_utils import to_categorical
import matplotlib.pyplot as plt
from keras.preprocessing.image import ImageDataGenerator
from sklearn.model_selection import train_test_split
from keras.callbacks import ReduceLROnPlateau
import sys

from keras.callbacks import TensorBoard
import numpy as np

input_ = Input(shape=(28,28,1,))

x = Conv2D(filters=32, kernel_size=(3, 3),activation='relu',kernel_initializer='he_normal',name='Conv_1')(input_)
x = Conv2D(32, kernel_size=(3, 3),activation='relu',kernel_initializer='he_normal',name='Conv_2')(x)
x = MaxPooling2D((2, 2),name='MaxPooling_1')(x)
x = Dropout(0.20,name='Dropout_1')(x)

x = Conv2D(64, (3, 3), activation='relu',padding='same',kernel_initializer='he_normal',name='Conv_3')(x)
x = Conv2D(64, (3, 3), activation='relu',padding='same',kernel_initializer='he_normal',name='Conv_4')(x)
x = MaxPooling2D(pool_size=(2, 2),name='MaxPooling_2')(x)
x = Dropout(0.25,name='Dropout_2')(x)

x = Conv2D(128, (3, 3), activation='relu',padding='same',kernel_initializer='he_normal',name='Conv_5')(x)
x = Dropout(0.25,name='Dropout_3')(x)

x = Flatten(name='Flatten')(x)
x = Dense(128, activation='relu',name='Dense_1')(x)
x= BatchNormalization(name='BatchNormalization')(x)
x = Dropout(0.25,name='Dropout_4')(x)
output= Dense(10,activation='softmax',name='Dense_2')(x)

model = Model(input_, output)
model.summary()
model.compile(optimizer='Adam',
              loss = 'categorical_crossentropy',
              metrics=['accuracy']
             )

datagen = ImageDataGenerator(
        featurewise_center=False,  # set input mean to 0 over the dataset
        samplewise_center=False,  # set each sample mean to 0
        featurewise_std_normalization=False,  # divide inputs by std of the dataset
        samplewise_std_normalization=False,  # divide each input by its std
        zca_whitening=False,  # apply ZCA whitening
        rotation_range=15, # randomly rotate images in the range (degrees, 0 to 180)
        zoom_range = 0.1, # Randomly zoom image 
        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
        height_shift_range=0.1,  # randomly shift images vertically (fraction of total height)
        horizontal_flip=False,  # randomly flip images
        vertical_flip=False)  # randomly flip images


learning_rate_reduction = ReduceLROnPlateau(monitor='val_acc', #监测指标
                                            patience=3, #3个epochs内指标不提升则更新学习率
                                            verbose=1, #显示输出
                                            factor=0.5, #lr = 0.5*lr
                                            min_lr=0.0001)#学习率的最小值

batch_size = 60
epochs = 30
(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train = np.expand_dims(x_train,axis=3) / 255.
x_test = np.expand_dims(x_test,axis=3) / 255.
y_train = to_categorical(y_train, num_classes = 10)#将标签转换成[0,1,0,0,0,0,0,0,0,0]
y_test = to_categorical(y_test, num_classes = 10)
x_train, x_val, y_train, y_val = train_test_split(x_train, y_train, test_size = 0.1, random_state=42)

tbCallBack = TensorBoard(log_dir="./model1",histogram_freq=5,write_grads=True)

datagen.fit(x_train)
history = model.fit_generator(datagen.flow(x_train,y_train, batch_size=batch_size),
                              epochs = epochs, validation_data = (x_val,y_val),
                              verbose = 1, steps_per_epoch=x_train.shape[0] // batch_size
                              , callbacks=[learning_rate_reduction,tbCallBack])


model.evaluate(x_test, y_test)