基于Keras：手写数字识别

一、概述

• 手写数字识别通常作为第一个深度学习在计算机视觉方面应用的示例，Mnist数据集在这当中也被广泛采用，可用于进行训练及模型性能测试；
• 模型的输入： 32*32的手写字体图片，这些手写字体包含0~9数字，也就是相当于10个类别的图片
• 模型的输出： 分类结果，0~9之间的一个数
• 下面通过多层感知器模型以及卷积神经网络的方式进行实现

二、基于多层感知器的手写数字识别

• 多层感知器的模型如下，其具有一层影藏层：

784个神经元	784个神经元	10个神经元
输入层	影藏层	输出层

• Mnist数据集此前可通过mnist.load_data()进行下载，但网址打不开，因此通过其他方式将数据集下载到本地，并在本地进行读取，数据集下载链接为：链接: https://pan.baidu.com/s/1ZlktkjqEGEJ0aZGQBQuqXg 提取码: br96
• 改编后的数据读取方式如下：

import numpy as np
def loadData(path="mnist.npz"):
    f = np.load(path)
    x_train, y_train = f['x_train'], f['y_train']
    x_test, y_test = f['x_test'], f['y_test']
    f.close()
    return (x_train, y_train), (x_test, y_test)

# 从Keras导入Mnist数据集
(x_train, y_train), (x_validation, y_validation) = loadData()

• 完整的实现代码如下：

import matplotlib.pyplot as plt
import numpy as np
from keras.models import Sequential
from keras.layers import Dense
from keras.utils import np_utils


def loadData(path="mnist.npz"):
    f = np.load(path)
    x_train, y_train = f['x_train'], f['y_train']
    x_test, y_test = f['x_test'], f['y_test']
    f.close()
    return (x_train, y_train), (x_test, y_test)


# 从Keras导入Mnist数据集
(x_train, y_train), (x_validation, y_validation) = loadData()

# 显示4张手写数字图片
plt.subplot(221)
plt.imshow(x_train[0], cmap=plt.get_cmap('gray'))

plt.subplot(222)
plt.imshow(x_train[1], cmap=plt.get_cmap('gray'))

plt.subplot(223)
plt.imshow(x_train[2], cmap=plt.get_cmap('gray'))

plt.subplot(224)
plt.imshow(x_train[3], cmap=plt.get_cmap('gray'))

plt.show()

# 设定随机种子
seed = 7
np.random.seed(seed)

num_pixels = x_train.shape[1] * x_train.shape[2]
print(num_pixels)

x_train = x_train.reshape(x_train.shape[0], num_pixels).astype('float32')
x_validation = x_validation.reshape(x_validation.shape[0], num_pixels).astype('float32')

# 格式化数据到0~1
x_train = x_train/255
x_validation = x_validation/255

# 进行one-hot编码
y_train = np_utils.to_categorical(y_train)
y_validation = np_utils.to_categorical(y_validation)
num_classes = y_validation.shape[1]
print(num_classes)


# 定义基准MLP模型
def create_model():
    model = Sequential()
    model.add(Dense(units=num_pixels, input_dim= num_pixels,kernel_initializer='normal', activation='relu'))
    model.add(Dense(units=num_classes, kernel_initializer='normal', activation='softmax'))

    # 编译模型
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
    return model


model = create_model()
model.fit(x_train, y_train, epochs=10, batch_size=200)

score = model.evaluate(x_validation, y_validation)
print('MLP: %.2f%%' % (score[1]*100))

• 程序运行结果如下

784
10
Epoch 1/10

  200/60000 [..............................] - ETA: 4:32 - loss: 2.3038 - acc: 0.1100
  600/60000 [..............................] - ETA: 1:37 - loss: 2.0529 - acc: 0.3283
 1000/60000 [..............................] - ETA: 1:02 - loss: 1.8041 - acc: 0.4710
 ...
 9472/10000 [===========================>..] - ETA: 0s
10000/10000 [==============================] - 1s 112us/step
MLP: 98.07%

三、基于卷积神经网络的手写数字识别

• 构建的卷积神经网络结构如下：

1 x 28 x 28个输入	32maps, 5 x 5	2 x 2	20%		128个	10个
输入层	卷积层	池化层	Dropout层	Flatten层	全连接层	输出层

Flatten层: Flatten层用来将输入“压平”，即把多维的输入一维化，常用在从卷积层到全连接层的过渡,举例如下

input size	---->>	output size
32 x 32 x 3	Flatten–>	3072

• 完整的实现代码如下：

import numpy as np
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import Dropout
from keras.layers import Flatten
from keras.layers.convolutional import Conv2D
from keras.layers.convolutional import MaxPooling2D
from keras.utils import np_utils
from keras import backend
backend.set_image_data_format('channels_first')


def loadData(path="mnist.npz"):
    f = np.load(path)
    x_train, y_train = f['x_train'], f['y_train']
    x_test, y_test = f['x_test'], f['y_test']
    f.close()
    return (x_train, y_train), (x_test, y_test)

# 从Keras导入Mnist数据集
(x_train, y_train), (x_validation, y_validation) = loadData()

# 设定随机种子
seed = 7
np.random.seed(seed)

x_train = x_train.reshape(x_train.shape[0], 1, 28, 28).astype('float32')
x_validation = x_validation.reshape(x_validation.shape[0], 1, 28, 28).astype('float32')
# 格式化数据到0~1
x_train = x_train/255
x_validation = x_validation/255

# 进行one-hot编码
y_train = np_utils.to_categorical(y_train)
y_validation = np_utils.to_categorical(y_validation)

# 定义模型
def create_model():
    model = Sequential()
    model.add(Conv2D(32, (5, 5), input_shape=(1, 28, 28), activation='relu'))
    model.add(MaxPooling2D(pool_size=(2, 2)))
    model.add(Dropout(0.2))
    model.add(Flatten())

    model.add(Dense(units=128, activation='relu'))
    model.add(Dense(units=10, activation='softmax'))

    # 编译模型
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

    return model


model = create_model()
model.fit(x_train, y_train, epochs=10, batch_size=200, verbose=2)

score = model.evaluate(x_validation, y_validation, verbose=0)
print('CNN_Small: %.2f%%' % (score[1]*100))

• 运行结果如下（明显感觉到运行时间较长）：

Epoch 1/10
 - 165s - loss: 0.2226 - acc: 0.9367
Epoch 2/10
 - 163s - loss: 0.0713 - acc: 0.9785
Epoch 3/10
 - 165s - loss: 0.0512 - acc: 0.9841
Epoch 4/10
 - 165s - loss: 0.0391 - acc: 0.9880
Epoch 5/10
 - 166s - loss: 0.0325 - acc: 0.9900
Epoch 6/10
 - 162s - loss: 0.0268 - acc: 0.9917
Epoch 7/10
 - 164s - loss: 0.0221 - acc: 0.9928
Epoch 8/10
 - 161s - loss: 0.0190 - acc: 0.9943
Epoch 9/10
 - 162s - loss: 0.0156 - acc: 0.9950
Epoch 10/10
 - 162s - loss: 0.0143 - acc: 0.9959
CNN_Small: 98.87%