import numpy as np

 from keras.datasets import mnist

 from keras.models import Sequential, Model

 from keras.layers.core import Dense, Activation, Dropout

 from keras.utils import np_utils

 import matplotlib.pyplot as plt

 import matplotlib.image as processimage

 # Load mnist RAW dataset

 # 训练集28*28的图片X_train = (60000, 28, 28) 训练集标签Y_train = (60000,1)

 # 测试集图片X_test  = (10000, 28, 28) 测试集标签Y_test  = (10000,1)

 (X_train, Y_train), (X_test, Y_test) = mnist.load_data()

 print(X_train.shape, Y_train.shape)

 print(X_test.shape, Y_test.shape)

 '''

 第一步，准备数据

 '''

 # Prepare 准备数据

 # Reshape 60k个图片，每个28*28的图片，降维成一个784的一维数组

 X_train = X_train.reshape(60000, 784)  # 28*28 = 784

 X_test = X_test.reshape(10000, 784)

 # set type into float32 设置成浮点型，因为使用的是GPU，GPU可以加速运算浮点型

 # CPU使用int型计算会更快

 X_train = X_train.astype('float32')  # astype SET AS TYPE INTO

 X_test = X_test.astype('float32')

 # 归一化颜色

 X_train = X_train/255  # 除以255个颜色，X_train(0, 255)-->(0, 1) 更有利于浮点运算

 X_test = X_test/255

 '''

 第二步，给神经网络设置基本参数

 '''

 # Prepare basic setups

 batch_sizes = 4096  # 一次给神经网络注入多少数据，别超过6万，和GPU内存有关

 nb_class = 10  # 设置多少个分类

 nb_epochs = 10  # 60k数据训练20次，一般小数据10次就够了

 '''

 第三步，设置标签

 '''

 # Class vectors label(7) into [0,0,0,0,0,0,0,1,0,1]  把7设置成向量

 Y_test = np_utils.to_categorical(Y_test, nb_class)  # Label

 Y_train = np_utils.to_categorical(Y_train, nb_class)

 '''

 第四步，设置网络结构

 '''

 model = Sequential()  # 顺序搭建层

 # 1st layer

 model.add(Dense(512, input_shape=(784,)))  # Dense是输出给下一层, input_dim = 784 [X*784]

 model.add(Activation('relu'))  # tanh

 model.add(Dropout(0.2))  # overfitting

 # 2nd layer

 model.add(Dense(256))  # 256是因为上一层已经输出512了，所以不用标注输入

 model.add(Activation('relu'))

 model.add(Dropout(0.2))

 # 3rd layer

 model.add(Dense(10))

 model.add(Activation('softmax'))  # 根据10层输出，softmax做分类

 '''

 第五步，编译compile

 '''

 model.compile(

     loss='categorical_crossentropy',

     optimizer='rmsprop',

     metrics=['accuracy']

 )

 # 启动网络训练 Fire up

 Trainning = model.fit(

     X_train, Y_train,

     batch_size=batch_sizes,

     epochs=nb_epochs,

     validation_data=(X_test, Y_test)

 )

 # 以上就可运行

 '''

 最后，检查工作

 '''

 # Trainning.history  # 检查训练历史

 # Trainning.params  # 检查训练参数

 # 拉取test里的图

 testrun = X_test[9999].reshape(1, 784)

 testlabel = Y_test[9999]

 print('label:-->', testlabel)

 print(testrun.shape)

 plt.imshow(testrun.reshape([28, 28]))

 # 判断输出结果

 pred = model.predict(testrun)

 print(testrun)

 print('label of test same Y_test[9999]-->>', testlabel)

 print('预测结果-->>', pred)

 print([final.argmax() for final in pred])  # 找到pred数组中的最大值

 # 用自己的画的图28*28预测一下 (不太准，可以用卷积)

 # 可以用PS创建28*28像素的图，且是灰度，没有色彩

 target_img = processimage.imread('/.../picture.jpg')

 print(' before reshape:->>', target_img.shape)

 plt.imshow(target_img)

 target_img = target_img.reshape(1, 784)  # reshape

 print(' after reshape:->>', target_img.shape)

 target_img = np.array(target_img)  # img --> numpy array

 target_img = target_img.astype('float32')  # int --> float32

 target_img /= 255  # (0,255) --> (0,1)

 print(target_img)

 mypred = model.predict(target_img)

 print(mypred)

 print(myfinal.argmax() for myfinal in mypred)