原始论文中的网络结构如下图：

 

keras生成的网络结构如下图：

代码如下：

import numpy as npfrom keras.preprocessing import imagefrom keras.models import Sequentialfrom keras.layers import Dense, Dropout, Flatten, Activationfrom keras.layers import Conv2D, MaxPooling2Dfrom keras.utils.vis_utils import plot_modelfrom keras.utils import np_utils# 从文件夹图像与标签文件载入数据def create_x(filenum, file_dir):    train_x = []    for i in range(filenum):        img = image.load_img(file_dir + str(i) + ".bmp", target_size=(28, 28))        img = img.convert('L')        x = image.img_to_array(img)        train_x.append(x)    train_x = np.array(train_x)    train_x = train_x.astype('float32')    train_x /= 255    return train_xdef create_y(classes, filename):    train_y = []    file = open(filename, "r")    for line in file.readlines():        train_y.append(int(line))    file.close()    train_y = np.array(train_y).astype('float32')    train_y = np_utils.to_categorical(train_y, classes)    return train_yclasses = 10X_train = create_x(55000, './train/')X_test = create_x(10000, './test/')Y_train = create_y(classes, 'train.txt')Y_test = create_y(classes, 'test.txt')# 从网络下载的数据集直接解析数据'''from tensorflow.examples.tutorials.mnist import input_datamnist = input_data.read_data_sets("MNIST/", one_hot=True)X_train, Y_train = mnist.train.images, mnist.train.labelsX_test, Y_test = mnist.test.images, mnist.test.labelsX_train = X_train.astype('float32')X_test = X_test.astype('float32')print(X_train.shape, X_test.shape)X_train = X_train.reshape(X_train.shape[0], 28, 28, 1)X_test = X_test.reshape(X_test.shape[0], 28, 28, 1)print(X_train[0])'''model = Sequential()model.add(Conv2D(filters=6, kernel_size=(5, 5), padding='valid', input_shape=(28, 28, 1), activation='tanh')) #C1model.add(MaxPooling2D(pool_size=(2, 2)))　　　　#S2model.add(Conv2D(filters=16, kernel_size=(5, 5), padding='valid', activation='tanh'))　　#C3model.add(MaxPooling2D(pool_size=(2, 2)))　　　　#S4model.add(Flatten())model.add(Dense(120, activation='tanh'))　　　　#C5model.add(Dense(84, activation='tanh'))　　　　#F6model.add(Dense(10, activation='softmax'))　　#outputmodel.summary()model.compile(loss='categorical_crossentropy', optimizer='sgd', metrics=['accuracy'])history = model.fit(X_train, Y_train, batch_size=500, epochs=50, verbose=1, validation_data=(X_test, Y_test))score = model.evaluate(X_test, Y_test, verbose=0)test_result = model.predict(X_test)result = np.argmax(test_result, axis=1)print(result)print('Test score:', score[0])print('Test accuracy:', score[1])plot_model(model, to_file='model.png', show_shapes=True, show_layer_names=False)

50次迭代，识别率在97%左右：

 相关测试数据可以在这里到。