import numpy as np # 导入Keras库 from keras.models import Sequential # 导入Keras层 from keras.layers import Dense,Dropout,Activation # 导入Keras优化器 from keras.optimizers import SGD,Adam # 导入Keras数据集 from keras.utils import to_categorical from keras.datasets import mnist print(np.__version__) # 1.23.5 print(keras.__version__) # 2.13.1 # 加载数据函数，返回训练集和测试集 def load_data(): # 加载mnist数据集，获取训练集和测试集 (x_train,y_train),(x_test,y_test)=mnist.load_data() # 获取前10000个样本 number=10000 # 获取前10000个样本的训练集 x_train=x_train[0:number] # 获取前10000个样本的标签 y_train=y_train[0:number] # 将训练集的形状转换为(number,28*28) x_train=x_train.reshape(number,28*28) # 将测试集的形状转换为(number,28*28) x_test=x_test.reshape(x_test.shape[0],28*28) # 将训练集的类型转换为float32 x_train=x_train.astype('float32') # 将测试集的类型转换为float32 x_test=x_test.astype('float32') # 将标签转换为one-hot编码 y_train=to_categorical(y_train,10) y_test=to_categorical(y_test,10) # 获取训练集 x_train=x_train # 获取测试集 x_test=x_test # 将训练集的像素值转换为0-1 x_train=x_train/255 # 将测试集的像素值转换为0-1 x_test=x_test/255 # 返回训练集和测试集 return (x_train,y_train),(x_test,y_test) (x_train,y_train),(x_test,y_test)=load_data()
查看x_train的形状
print(x_train.shape) # (10000, 784)
表示有10000个数据，每个数据有784个像素点
可视化x_train
import matplotlib.pyplot as plt plt.imshow(x_train[0].reshape(28,28),cmap='gray') plt.show()
查看y_train的形状
print(y_train.shape) # (10000, 10)
查看y_train
y_train[0] array([0., 0., 0., 0., 0., 1., 0., 0., 0., 0.], dtype=float32)
# 定义一个模型 model = Sequential() # 添加一个输入层，输入维度为28*28，输出维度为633，激活函数为sigmoid model.add(Dense(input_dim=28*28, units=633,activation='sigmoid')) # 添加一个隐藏层，输出维度为633，激活函数为sigmoid model.add(Dense(units=633,activation="sigmoid")) # 添加一个隐藏层，输出维度为633，激活函数为sigmoid model.add(Dense(units=633,activation="sigmoid")) # 添加一个输出层，输出维度为10，激活函数为softmax model.add(Dense(units=10,activation="softmax"))
# 编译模型，损失函数为mse，优化器为SGD，目标函数为accuracy model.compile(loss="mse",optimizer=SGD(learning_rate=0.1), metrics=['accuracy'])
# 训练模型，每次训练100个样本，训练20次 model.fit(x_train, y_train, batch_size=100, epochs=20) # 评估模型，计算测试集上的准确率 restult = model.evaluate(x_test, y_test) print("Test Acc:",restult[1])
分类问题mse不适合，将loss mse function 改为categorical_crossentropy，看看有怎样的差别
model.compile(loss="categorical_crossentropy",optimizer=SGD(learning_rate=0.1), metrics=['accuracy'])
当我们一换categorical_crossentropy，在train set上的结果就起飞了。得到87.34%的正确率，现在就比较有train起来了。
Epoch 20/20 100/100 [==============================] - 1s 14ms/step - loss: 0.4975 - accuracy: 0.8551 1/1 [==============================] - 0s 426ms/step - loss: 0.4771 - accuracy: 0.8661 Train Acc: 0.866100013256073 313/313 [==============================] - 1s 3ms/step - loss: 0.5052 - accuracy: 0.8471 Test Acc: 0.847100019454956