参考博客：http://blog.csdn.net/gjq246/article/details/71787217?ABstrategy=codes_snippets_optimize_v3

安装模型图片导出模块

sudo pip install pydot

sudo pip install graphviz

sudo pip install pydot-ng

sudo apt-get install graphviz


安装h5py的命令如下（模型保存模块）：
sudo pip install cython
sudo apt-get install libhdf5-dev
sudo pip install h5py

记录一下代码：

# -*- coding: UTF-8 -*-

# mnist神经网络训练，采用LeNet-5模型

import os
import cv2
import numpy as np

from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Flatten
from keras.layers.core import Dense, Dropout, Activation, Flatten
from keras.layers.advanced_activations import PReLU
from keras.optimizers import SGD, Adadelta, Adagrad

from keras.utils import np_utils
from keras.utils.vis_utils import plot_model

import h5py
from keras.models import model_from_json


def loadData(path, number):
    data = np.empty((number, 1, 28, 28), dtype="float32")  # empty与ones差不多原理，但是数值随机，类型随后面设定
    labels = np.empty((number,), dtype="uint8")
    listImg = os.listdir(path)
    count = 0
    for img in listImg:
        imgData = cv2.imread(path + '/' + img, 0)  # 数据
        l = int(img.split('-')[0])  # 答案
        arr = np.asarray(imgData, dtype="float32")  # 将img数据转化为数组形式
        data[count, :, :, :] = arr  # 将每个三维数组赋给data
        labels[count] = l  # 取该图像的数值属性作为标签
        count = count + 1
        print path, " loaded ", count
        if count >= number:
            break
    return data, labels


# 从图片文件加载数据
# the data, shuffled and split between train and test sets
(trainData, trainLabels), (testData, testLabels) = mnist.load_data()

# 训练数据 60000张手写图片，28*28*1
# 测试数据 10000张手写图片，28*28*1

trainData = trainData.reshape(60000, 784)
testData = testData.reshape(10000, 784)

trainLabels = np_utils.to_categorical(trainLabels, 10)
# label为0~9共10个类别，keras要求格式为binary class matrices,转化一下，直接调用keras提供的这个函数
testLabels = np_utils.to_categorical(testLabels, 10)

# tf或th为后端，采取不同参数顺序
# th
# if K.image_data_format() == 'channels_first':
# -x_train.shape[0]=6000
# x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
# -x_train.shape:(60000, 1, 28, 28)
# x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
# x_test.shape:(10000, 1, 28, 28)
# 单通道灰度图像,channel=1
# input_shape = (1, img_rows, img_cols)
# else: #tf
# x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
# x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
# input_shape = (img_rows, img_cols, 1)

# tensorflow后端
trainData = trainData.reshape(trainData.shape[0], 28, 28, 1)
testData = testData.reshape(testData.shape[0], 28, 28, 1)

# 建立一个Sequential模型
model = Sequential()

# model.add(Conv2D(4, 5, 5, border_mode='valid',input_shape=(28,28,1)))
# 第一个卷积层，4个卷积核，每个卷积核5*5,卷积后24*24，第一个卷积核要申明input_shape(通道，大小) ,激活函数采用“tanh”
model.add(Conv2D(filters=4, kernel_size=(5, 5), padding='valid', input_shape=(28, 28, 1), activation='tanh'))

# model.add(Conv2D(8, 3, 3, subsample=(2,2), border_mode='valid'))
# 第二个卷积层，8个卷积核，不需要申明上一个卷积留下来的特征map，会自动识别，下采样层为2*2,卷完且采样后是11*11
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(filters=8, kernel_size=(3, 3), padding='valid', activation='tanh'))
# model.add(Activation('tanh'))

# model.add(Conv2D(16, 3, 3, subsample=(2,2), border_mode='valid'))
# 第三个卷积层，16个卷积核，下采样层为2*2,卷完采样后是4*4
model.add(Conv2D(filters=16, kernel_size=(3, 3), padding='valid', activation='tanh'))
model.add(MaxPooling2D(pool_size=(2, 2)))
# model.add(Activation('tanh'))

model.add(Flatten())
# 把多维的模型压平为一维的，用在卷积层到全连接层的过度

# model.add(Dense(128, input_dim=(16*4*4), init='normal'))
# 全连接层，首层的需要指定输入维度16*4*4,128是输出维度，默认放第一位
model.add(Dense(128, activation='tanh'))

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

# model.add(Dense(10, input_dim= 128, init='normal'))
# 第二层全连接层，其实不需要指定输入维度，输出为10维，因为是10类
model.add(Dense(10, activation='softmax'))
# model.add(Activation('softmax'))
# 激活函数“softmax”，用于分类

# 训练CNN模型

sgd = SGD(lr=0.05, momentum=0.9, decay=1e-6, nesterov=True)
# 采用随机梯度下降法，学习率初始值0.05,动量参数为0.***习率衰减值为1e-6,确定使用Nesterov动量
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
# 配置模型学习过程，目标函数为categorical_crossentropy：亦称作多类的对数损失，注意使用该目标函数时，需要将标签转化为形如(nb_samples, nb_classes)的二值序列，第18行已转化，优化器为sgd

model.fit(trainData, trainLabels, batch_size=100, epochs=20, shuffle=True, verbose=1, validation_split=0.2)
# 训练模型，训练nb_epoch次，bctch_size为梯度下降时每个batch包含的样本数，验证集比例0.2,verbose为显示日志，shuffle是否打乱输入样本的顺序

# 输出模型图片
plot_model(model, to_file='model2.png', show_shapes=True, show_layer_names=False)

print model.metrics_names
# 对测试数据进行测试
print model.evaluate(testData, testLabels,
                     verbose=0,
                     batch_size=500);

# 保存model
json_string = model.to_json()
open('my_model_architecture.json', 'w').write(json_string)
model.save_weights('my_model_weights.h5')