深度学习实战之MNIST

既是实战，也是入门~

MNIST介绍

MNIST是深度学习领域的一个经典数据集，内含60000张训练图像与10000张预测图像，每张图片为28像素*28像素的灰度图像，并被划分到10个类别中(0-9)。

MNIST手写数字识别，正是深度学习里的Hello World。

加载数据集

mnist数据预加载在keras库中，其中包括4个numpy数组

from keras.datasets import mnist
(train_images,train_labels),(test_images,test_labels) = mnist.load_data()
print(train_images.shape,test_images.shape)
print(train_labels[:20])

构建网络

网络共2个dense层（即全连接层）

第一层网络共512个隐藏单元（hidden unit），激活函数为relu

第一层网络共10个隐藏单元，激活函数为softmax

关于指定输入数据的shape，可以看http://ducknew.cf/posts/e9e6cec8/

from keras import models,layers
network=models.Sequential()
network.add(layers.Dense(512,activation='relu',input_shape=(28*28,)))
network.add(layers.Dense(10,activation='softmax'))

optimizer（优化器）：该参数可指定为已预定义的优化器名，如 rmsprop 、 adagrad ，或一 个 Optimizer 类的对象
loss（损失函数）：该参数为模型试图最小化的目标函数，它可为预定义的损失函数名， 如 categorical_crossentropy 、 mse
metrics（指标列表）：对分类问题，我们一般将该列表设置为 metrics=[‘accuracy’] 。指标可以是一个预 定义指标的名字,也可以是一个用户定制的函数

network.compile(optimizer='rmsprop',
                loss='categorical_crossentropy',
                metrics=['accuracy'])

数据预处理

原数据shape为(60000,28,28)，类型为uint8，取值范围为【0，255】

转换后数据shape为(60000,28*28)，类型为float32，取值范围为【0，1】

train_images=train_images.reshape((60000,28*28))
train_images = train_images.astype('float32') / 255
test_images=test_images.reshape((10000,28*28))
test_images = test_images.astype('float32') / 255

准备标签

对标签进行分类编码

from tensorflow.keras.utils import to_categorical
train_labels = to_categorical(train_labels)
test_labels = to_categorical(test_labels)

训练模型

epochs被定义为向前和向后传播中所有批次的单次训练迭代

举个例子

训练集有1000个样本，batchsize=10，那幺：

训练完整个样本集需要：

100次iteration，1次epoch

oneepoch
= numbers ofiterations
= N = 训练样本的数量/batch_size

batch_size可以看这里: http://ducknew.cf/posts/e9e6cec8/

network.fit(train_images,train_labels,epochs=10,batch_size=128)

运行过程显示：

Epoch 1/10
469/469 [==============================] - 15s 6ms/step - loss: 0.4178 - accuracy: 0.8784
Epoch 2/10
469/469 [==============================] - 3s 5ms/step - loss: 0.1119 - accuracy: 0.9669
Epoch 3/10
469/469 [==============================] - 2s 5ms/step - loss: 0.0711 - accuracy: 0.9784
Epoch 4/10
469/469 [==============================] - 2s 5ms/step - loss: 0.0504 - accuracy: 0.9852
Epoch 5/10
469/469 [==============================] - 3s 5ms/step - loss: 0.0376 - accuracy: 0.9888
Epoch 6/10
469/469 [==============================] - 2s 5ms/step - loss: 0.0263 - accuracy: 0.9923
Epoch 7/10
469/469 [==============================] - 2s 5ms/step - loss: 0.0201 - accuracy: 0.9942
Epoch 8/10
469/469 [==============================] - 2s 5ms/step - loss: 0.0142 - accuracy: 0.9959
Epoch 9/10
469/469 [==============================] - 3s 5ms/step - loss: 0.0116 - accuracy: 0.9968
Epoch 10/10
469/469 [==============================] - 2s 5ms/step - loss: 0.0094 - accuracy: 0.9977

评估模型

loss是网络在测试数据上的损失，acc是网络在测试数据上的精度

test_loss,test_acc = network.evaluate(test_images,test_labels)
print(f'test_loss: {test_loss}, test_acc: {test_acc}')

结果：

test_loss: 0.0742919072508812, test_acc: 0.9818999767303467

可以发现test_acc
<训练过程中的accuracy
，这种训练精度高而测试精度低的情况一般是**过拟合（overfit）**造成的。

CNN处理MNIST

网络结构

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 26, 26, 32)        320       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 13, 13, 32)        0         
_________________________________________________________________
dropout (Dropout)            (None, 13, 13, 32)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 11, 11, 64)        18496     
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 5, 5, 64)          0         
_________________________________________________________________
flatten (Flatten)            (None, 1600)              0         
_________________________________________________________________
dense (Dense)                (None, 64)                102464    
_________________________________________________________________
dropout_1 (Dropout)          (None, 64)                0         
_________________________________________________________________
dense_1 (Dense)              (None, 10)                650       
=================================================================
Total params: 121,930
Trainable params: 121,930
Non-trainable params: 0

网络代码

def build_model():
    model = models.Sequential()
    model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28,28,1)))
    model.add(layers.MaxPooling2D((2, 2)))
    model.add(layers.Dropout(0.25))
    model.add(layers.Conv2D(64, (3, 3), activation='relu'))
    model.add(layers.MaxPooling2D((2, 2)))
    model.add(layers.Flatten())
    model.add(layers.Dense(64, activation='relu'))
    model.add(layers.Dropout(0.5))
    model.add(layers.Dense(10, activation='softmax'))
    # print(model.summary())
    model.compile(optimizer='rmsprop',
                    loss='categorical_crossentropy',
                    metrics=['accuracy'])
    return model

loss与acc曲线

history=model.fit(train_images, train_labels, epochs=15, batch_size=128)
    print(history.history.keys())
    
    acc = history.history['accuracy']
    loss = history.history['loss']
    epochs = range(1, len(acc) + 1)
    
    plt.plot(epochs, acc, 'bo', label='Training accuracy')
    plt.title('Training  and validation accuracy')
    plt.legend()
    plt.figure()
    
    plt.plot(epochs, loss, 'bo', label='Training loss')
    plt.title('Training  and validation loss')
    plt.legend()
    plt.show()

事实上，卷积神经网络往往能比全连接网络取得更高的预测的精度。