基于mnist数据集，建立mlp模型，实现0-9数字的十分类

mnist数据集是机器学习领域中非常经典的一个数据集，由60000个训练样本和10000个测试样本组成，每个样本都是一张28*28像素的灰度手写数字图片。一共四个文件，训练集、训练集标签、测试集、测试集标签

• 实现mnist数据载入，可视化图形数字
• 实现数据预处理，图像数据维度转换与归一化，输出结果格式转换
• 计算模型在预测数据集的准确率
• 模型结构：两层隐藏层，每层有392个神经元

实现mnist数据载入，可视化图形数字

• 加载mnist数据集

  from matplotlib import pyplot as plt
  
  (X_train, y_train), (X_test, y_test) = mnist.load_data()
  print(type(X_train), X_train.shape)
  

  结果是 (60000, 28, 28)

• 可视化第一个图形数字

from matplotlib import pyplot as plt
# 可视化部分数据
img1 = X_train[0]
fig1 = plt.figure(figsize=(5, 5))
plt.imshow(img1)
plt.title(y_train[0])
plt.show()

结果是

从图形中可以看出来大概是数字5输出维度print(img1.shape),结果的确是(28, 28)每个样本都是一张28*28像素的灰度手写数字图片

实现数据预处理，图像数据维度转换与归一化，输出结果格式转换

• 图像数据维度转换

  要把28*28=784的像素灰度转换成784列

  feature_size = img1.shape[0] * img1.shape[1]
  X_train_format = X_train.reshape(X_train.shape[0], feature_size)
  print(X_train_format.shape)
  结果是(60000, 784)
  

• 归一化

  为了更快的进行mlp的迭代和求解

  X_train_normal = X_train_format / 255
  X_test_normal = X_test_format / 255
  

• 输出结果格式转换

  from keras.utils import to_categorical
  y_train_format = to_categorical(y_train)
  y_test_format = to_categorical(y_test)
  print(y_train[0]) 
  print(y_train_format[0])
  输出结果是5和[0. 0. 0. 0. 0. 1. 0. 0. 0. 0.]
  

  模型结构搭建

• 建立模型结构

  from keras.models import Sequential
  from keras.layers import Dense, Activation
  mlp = Sequential()
  mlp.add(Dense(units=392, activation='sigmoid', input_dim=feature_size))
  mlp.add(Dense(units=392, activation='sigmoid'))
  mlp.add(Dense(units=10, activation='softmax'))
  mlp.summary()
  

  输出结果是

• 模型配置

  # 模型配置
  mlp.compile(loss='categorical_crossentropy', optimizer='adam')
  

• 模型训练

  # 模型训练
  mlp.fit(X_train_normal, y_train_format, epochs=10)
  

• 模型评估

  # 模型评估
  # 训练集训练准确率
  y_train_predict = mlp.predict_classes(X_train_normal)
  print(y_train_predict)
  accuracy_train = accuracy_score(y_train, y_train_predict)
  print(accuracy_train)
  # 测试集训练准确率
  y_test_predict = mlp.predict_classes(X_test_normal)
  print(y_test_predict)
  accuracy_test = accuracy_score(y_test, y_test_predict)
  print(accuracy_test)
  

  输出结果
  

可以看到这个模型训练的很好，训练集的准确率是0.99，测试集的准确率是0.98

• 直观查看模型的预测结果

  # 测试集图形的可视化展示
  img2 = X_test[100]
  fig2 = plt.figure(figsize=(3, 3))
  plt.imshow(img2)
  plt.title(y_test_predict[100])
  plt.show()
  

结果是

可以看到，预测的结果和照片中是一样的。

全部代码为

# 加载mnist数据集

from keras.datasets import mnist
from keras.utils import to_categorical
from keras.models import Sequential
from keras.layers import Dense, Activation
from sklearn.metrics import accuracy_score
from matplotlib import pyplot as plt

(X_train, y_train), (X_test, y_test) = mnist.load_data()
print(type(X_train), X_train.shape)

# 可视化部分数据
img1 = X_train[0]
fig1 = plt.figure(figsize=(3, 3))
plt.imshow(img1)
plt.title(y_train[0])
plt.show()

feature_size = img1.shape[0] * img1.shape[1]
X_train_format = X_train.reshape(X_train.shape[0], feature_size)
X_test_format = X_test.reshape(X_test.shape[0], feature_size)
print(X_test_format.shape)

X_train_normal = X_train_format / 255
X_test_normal = X_test_format / 255

y_train_format = to_categorical(y_train)
y_test_format = to_categorical(y_test)
print(y_train[0])
print(y_train_format[0])

# 建立模型结构
mlp = Sequential()
mlp.add(Dense(units=392, activation='sigmoid', input_dim=feature_size))
mlp.add(Dense(units=392, activation='sigmoid'))
mlp.add(Dense(units=10, activation='softmax'))
mlp.summary()

# 模型配置
mlp.compile(loss='categorical_crossentropy', optimizer='adam')
# 模型训练
mlp.fit(X_train_normal, y_train_format, epochs=10)
# 模型评估
# 训练集训练准确率
y_train_predict = mlp.predict_classes(X_train_normal)
print(y_train_predict)
accuracy_train = accuracy_score(y_train, y_train_predict)
print(accuracy_train)
# 测试集训练准确率
y_test_predict = mlp.predict_classes(X_test_normal)
print(y_test_predict)
accuracy_test = accuracy_score(y_test, y_test_predict)
print(accuracy_test)

# 测试集图形的可视化展示
img2 = X_test[100]
fig2 = plt.figure(figsize=(3, 3))
plt.imshow(img2)
plt.title(y_test_predict[100])
plt.show()
 = accuracy_score(y_test, y_test_predict)
print(accuracy_test)

# 测试集图形的可视化展示
img2 = X_test[100]
fig2 = plt.figure(figsize=(3, 3))
plt.imshow(img2)
plt.title(y_test_predict[100])
plt.show()