Keras Mnist

在这里，我们将利用Keras搭建一个深度学习网络对mnist数据集进行识别。

• 本文参考 keras-mnist-tutorial
• 整个代码分为三个部分：
  
  1. 数据准备
  2. 模型搭建
  3. 训练优化

让我们开始吧

首先先导入一些模块

%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt

from keras.datasets import mnist
from keras.models import Sequential
from keras.layers.core import Dense, Activation, Dropout
from keras.utils import np_utils

数据准备

我们通过keras自带的数据集mnist进行导入数据，然后对其归一化处理，并且将原二维数据变成一维数据，作为网络的输入。

读入mnist数据集。可以看到每条样本是一个28*28的矩阵，共有60000个训练数据，10000个测试数据。

(X_train, y_train), (X_test, y_test) = mnist.load_data();
print(x_train.shape)
print(x_test.shape)

(60000, 28, 28)
(10000, 28, 28)

将一些样本图像打印出来看看

for i in range():
    plt.subplot(,,i+)
    plt.imshow(X_train[i], cmap='gray', interpolation='none')
    plt.title("Class {}".format(y_train[i]))

将二维数据变成一维数据

X_train = X_train.reshape(len(X_train), -)
X_test = X_test.reshape(len(X_test), -)

接下来对数据进行归一化。原来的数据范围是[0,255]，我们通过归一化时靠近0附近。归一化的方式有很多，大家随意。

# uint不能有负数，我们先转为float类型
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train = (X_train - ) / 
X_test = (X_test - ) / 

接下来 One-hot encoding

nb_classes = 
y_train = np_utils.to_categorical(y_train, nb_classes)
y_test = np_utils.to_categorical(y_test, nb_classes)

搭建网络

数据已经准备好了，接下来我们进行网络的搭建，我们的网络有三层，都是全连接网络，大概长的像这样

这里新遇到一个Dropout，这是一种防止过拟合(overfitting)的方法，详见Dropout层

model = Sequential()

model.add(Dense(, input_shape=(,), kernel_initializer='he_normal'))
model.add(Activation('relu'))
model.add(Dropout()) 

model.add(Dense(, kernel_initializer='he_normal'))
model.add(Activation('relu'))
model.add(Dropout()) 

model.add(Dense(nb_classes))
model.add(Activation('softmax'))

OK！模型搭建好了，我们通过编译对学习过程进行配置

model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

那么我们进行训练吧

model.fit(X_train, y_train, epochs=, batch_size=, verbose=, validation_split=)

Train on 57000 samples, validate on 3000 samples
Epoch 1/20
57000/57000 [==============================] - 19s 327us/step - loss: 0.0811 - acc: 0.9761 - val_loss: 0.0702 - val_acc: 0.9810
Epoch 2/20
57000/57000 [==============================] - 19s 328us/step - loss: 0.0752 - acc: 0.9772 - val_loss: 0.0720 - val_acc: 0.9813
Epoch 3/20
57000/57000 [==============================] - 19s 331us/step - loss: 0.0687 - acc: 0.9788 - val_loss: 0.0670 - val_acc: 0.9830
Epoch 4/20
57000/57000 [==============================] - 20s 350us/step - loss: 0.0667 - acc: 0.9794 - val_loss: 0.0755 - val_acc: 0.9810
Epoch 5/20
57000/57000 [==============================] - 20s 353us/step - loss: 0.0688 - acc: 0.9794 - val_loss: 0.0671 - val_acc: 0.9820
Epoch 6/20
57000/57000 [==============================] - 20s 346us/step - loss: 0.0639 - acc: 0.9807 - val_loss: 0.0744 - val_acc: 0.9790
Epoch 7/20
57000/57000 [==============================] - 20s 342us/step - loss: 0.0626 - acc: 0.9805 - val_loss: 0.0685 - val_acc: 0.9837
Epoch 8/20
57000/57000 [==============================] - 21s 365us/step - loss: 0.0669 - acc: 0.9796 - val_loss: 0.0988 - val_acc: 0.9757
Epoch 9/20
57000/57000 [==============================] - 20s 345us/step - loss: 0.0605 - acc: 0.9819 - val_loss: 0.0769 - val_acc: 0.9833
Epoch 10/20
57000/57000 [==============================] - 19s 338us/step - loss: 0.0592 - acc: 0.9820 - val_loss: 0.0576 - val_acc: 0.9870
Epoch 11/20
57000/57000 [==============================] - 19s 336us/step - loss: 0.0600 - acc: 0.9822 - val_loss: 0.0689 - val_acc: 0.9847
Epoch 12/20
57000/57000 [==============================] - 20s 345us/step - loss: 0.0625 - acc: 0.9813 - val_loss: 0.0689 - val_acc: 0.9843
Epoch 13/20
57000/57000 [==============================] - 20s 346us/step - loss: 0.0573 - acc: 0.9829 - val_loss: 0.0679 - val_acc: 0.9853
Epoch 14/20
57000/57000 [==============================] - 19s 342us/step - loss: 0.0555 - acc: 0.9833 - val_loss: 0.0642 - val_acc: 0.9850
Epoch 15/20
57000/57000 [==============================] - 20s 359us/step - loss: 0.0571 - acc: 0.9831 - val_loss: 0.0779 - val_acc: 0.9833
Epoch 16/20
57000/57000 [==============================] - 21s 361us/step - loss: 0.0564 - acc: 0.9831 - val_loss: 0.0610 - val_acc: 0.9867
Epoch 17/20
57000/57000 [==============================] - 20s 354us/step - loss: 0.0574 - acc: 0.9834 - val_loss: 0.0669 - val_acc: 0.9867
Epoch 18/20
57000/57000 [==============================] - 20s 353us/step - loss: 0.0526 - acc: 0.9848 - val_loss: 0.0863 - val_acc: 0.9830
Epoch 19/20
57000/57000 [==============================] - 20s 349us/step - loss: 0.0548 - acc: 0.9832 - val_loss: 0.0726 - val_acc: 0.9847
Epoch 20/20
57000/57000 [==============================] - 20s 352us/step - loss: 0.0512 - acc: 0.9845 - val_loss: 0.0735 - val_acc: 0.9860





<keras.callbacks.History at 0x2904822cd30>

训练完毕，测试测试

loss, accuracy = model.evaluate(X_test, y_test)
print('Test loss:', loss)
print('Accuracy:', accuracy)

10000/10000 [==============================] - 1s 80us/step
Test loss: 0.0864374790877
Accuracy: 0.9817