第五章 深度学习用于计算机视觉

""" 
1.卷积神经网络CNN--手写数字分类-MNIST数据集 
"""
# 1.网络架构 实例化一个小型的卷积神经网络
from keras import models
from keras import layers

network = models.Sequential()     # 因为需要将同一个模型多次实例化，所以用一个函数Sequential来构建模型，构造器
network.add(layers.Conv2D(32,(3,3), activation='relu', input_shape=(28,28,1)))   # 输入3D
network.add(layers.MaxPooling2D((2,2)))
network.add(layers.Conv2D(64,(3,3), activation='relu'))     # 卷积核大小(3,3)
network.add(layers.MaxPooling2D((2,2)))                     # 池化窗口(2,2)
network.add(layers.Conv2D(64,(3,3), activation='relu'))
network.summary() 

# 2.在卷积神经网络上添加分类器
network.add(layers.Flatten())                    # 3D变成1D 展平输出(3, 3, 64)
network.add(layers.Dense(64, activation='relu'))
network.add(layers.Dense(10, activation='softmax'))   # 10个输出
network.summary()                 # 查看网络结构

# 3.在MNIST图像上训练卷积神经网络
from keras.datasets import mnist           # 加载Keras中的MNIST数据集
from keras.utils import to_categorical     # 将data列表转化成为整数张量

(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

# 4.预处理准备图像数据
train_images = train_images.reshape((60000, 28,28, 1))   # 4D
train_images = train_images.astype('float32')/255        # 预处理，取值范围0～1

test_images = test_images.reshape(10000, 28,28, 1)
test_images = test_images.astype('float32')/255

train_labels = to_categorical(train_labels)     # 准备标签：将数据列表转换为整数张量
test_labels = to_categorical(test_labels)


network.compile(optimizer='rmsprop',                # rmsprop优化器
                loss='categorical_crossentropy',    # 损失函数：categorical_crossentropy：分类交叉熵
                metrics=['accuracy'])               # 指标：精度

network.fit(train_images, train_labels, epochs=5, batch_size=64)      # 用训练集训练模型
test_loss, test_acc = network.evaluate(test_images, test_labels)      # 评估测试集

print('test_loss:', test_loss)
print('test_acc:', test_acc)


print(train_images.shape)   # train_images 4D (60000, 28, 28, 1)
print(len(train_images))    # 60000
print(train_labels)        # 稀疏矩阵
print(train_images.ndim)   # 轴的个数   4D
print(train_images.dtype)  # 数据类型   float32

代码运行结果：99%