keras实现mnist识别

# 包引入
import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Flatten, Dense
import matplotlib.pyplot as plt

# 1.下载数据集
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

# 案例图像显示
plt.figure(figsize=(20,10))  # 绘制(20,10)图像
for i in range(20):
    plt.subplot(5,10,i+1)  # 5*10子图
    plt.xticks([])  # x,y去标签
    plt.yticks([])
    plt.grid(False)
    plt.imshow(train_images[i], cmap=plt.cm.binary)
    plt.xlabel(train_labels[i])
plt.show()

# 归一化
train_images, test_images = train_images / 255.0, test_images / 255.0
# 调整格式
train_images = train_images.reshape((60000, 28, 28, 1))
test_images = test_images.reshape((10000, 28, 28, 1))

# CNN建模
model = Sequential()
# 第一层
# 卷积层1: 卷积核3*3, 共32个
model.add(Conv2D(32,(3,3), activation='relu', input_shape=(28,28,1)))
# 池化层1: 采样2*2
model.add(MaxPooling2D((2,2)))

# 第二层
# 卷积层2: 卷积核3*3, 共64个
model.add(Conv2D(64,(3,3), activation='relu'))
# 池化层1: 采样2*2
model.add(MaxPooling2D((2, 2)))

# Flatten层: 连接卷积层与全连接层
model.add(Flatten())
# 全连接层
model.add(Dense(64, activation='relu'))
# 输出层
model.add(Dense(10))

# 打印网络结构
model.summary()

# 编译模型
# 优化器、损失函数、矩阵
model.compile(optimizer='adam',
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

# 训练模型
history = model.fit(train_images, train_labels, epochs=10,
                    validation_data=(test_images, test_labels))

# 预测
pre_outputs = pre = model.predict(test_images)

活动地址：CSDN21天学习挑战赛