TensorFlow学习总结（五）实现MNIST分类及最佳实践样例

一、MNIST认知(def)

（1） MNIST数据集中的每一张图片都代表了0~9中的一个数，即10类。

（2）图片大小都是28×28=784

（3）下载地址：http://yann.lecun.com/exdb/mnist

（4）MNIST数据集划分为train(55000)、validation(5000)、test(10000)三个数据集

二、tensorflow实现及详细分析理解

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data

'''定义参数'''
INPUT_NODE = 784                    #输入层节点数；即图片像素28×28=784
OUTPUT_NODE = 10                    #输出层节点数；即类别数0~9共10类
LAYER1_NODE = 500                   #隐藏层节点数；
BATCH_SIZE = 100                    #一个训练batch中的训练个数
LEARNING_RATE_BASE = 0.8            #基础学习率
LEARNING_RATE_DECAY = 0.99          #学习率的衰减率
REGULARIZATION_RATE = 0.0001        #描述模型复杂度的正则化项在损失函数中的系数；即lambda
TRAINING_STEPS = 30000              #训练轮数
MOVING_AVERAGE_DECAY = 0.99         #滑动平均衰减率


'''前向传播网络 ：包含输入、滑动平均值形参、所有参数'''
def inference(input_tensor, avg_class, weights1, biases1, weights2, biases2):
    if avg_class == None:           #无滑动平均
        layer1 = tf.nn.relu(tf.matmul(input_tensor, weights1) + biases1)
        return tf.matmul(layer1, weights2) + biases2
    else:                           #有滑动平均
        layer1 = tf.nn.relu(tf.matmul(input_tensor, avg_class.average(weights1)) + avg_class.average(biases1))
        return tf.matmul(layer1, avg_class.average(weights2)) + avg_class.average(biases2)

'''训练模型'''
def train(mnist):
    x = tf.placeholder(tf.float32, [None, INPUT_NODE], name='x-input')                              #输入占位
    y_ = tf.placeholder(tf.float32, [None, OUTPUT_NODE], name='y-input')                            #输出占位；真实值
    weights1 = tf.Variable(tf.truncated_normal([INPUT_NODE, LAYER1_NODE], stddev=0.1))              #隐藏层参数  #权重1
    biases1 = tf.Variable(tf.constant(0.1, shape=[LAYER1_NODE]))                                                 #偏置1
    weights2 = tf.Variable(tf.truncated_normal([LAYER1_NODE, OUTPUT_NODE], stddev=0.1))             #输出层参数  #权重2
    biases2 = tf.Variable(tf.constant(0.1, shape=[OUTPUT_NODE]))                                                 #偏置2
    y = inference(x, None, weights1, biases1, weights2, biases2)                                    #前向传播结果；预测
    global_step = tf.Variable(0, trainable=False)                                                   #当前轮数
    variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)        #初始化滑动平均类
    variables_averages_op = variable_averages.apply(tf.trainable_variables())                       #在所有神经网络参数上进行滑动平均
    average_y = inference(x, variable_averages, weights1, biases1, weights2, biases2)               #滑动平均之后的预测值
    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=tf.argmax(y_, 1), logits=y)#交叉熵
    cross_entropy_mean = tf.reduce_mean(cross_entropy)                                              #当前batch中交叉熵平均值
    regularizer = tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)                             #正则化损失函数初始化
    regularization = regularizer(weights1) + regularizer(weights2)                                  #计算模型的正则化损失（一般只计算权重，不算偏置）
    loss = cross_entropy_mean + regularization                                                      #总损失函数；交叉熵+正则化
    learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE, global_step, mnist.train.num_examples / BATCH_SIZE, LEARNING_RATE_DECAY)#学习率初始化
    train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)#整体优化损失函数；反向传播网络
    with tf.control_dependencies([train_step, variables_averages_op]):                              #同时更新参数和对应的滑动平均值
        train_op = tf.no_op(name='train')
    correct_prediction = tf.equal(tf.argmax(average_y, 1), tf.argmax(y_, 1))                        #检验前向传播网络滑动平均值是否正确；1代表每一行选取最大值对应的下标
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))                              #模型在这组数据上的正确率

    '''开始训练'''
    with tf.Session() as sess:
        tf.initialize_all_variables().run()                                                         #变量初始化
        validate_feed = {x: mnist.validation.images,                                                #输入验证数据和答案
                         y_: mnist.validation.labels}
        test_feed = {x:mnist.test.images, y_: mnist.test.labels}                                    #输入测试数据和答案
        for i in range(TRAINING_STEPS):                                                             #循环TRAINING_STEPS次
             if i % 1000 == 0:                                                                          #每一千次打印一次
                 validate_acc = sess.run(accuracy, feed_dict=validate_feed)                             #验证机准确率
                 print("After %d training step(s), validation accuracy"
                       "using average model is %g" % (i, validate_acc))
             xs, ys = mnist.train.next_batch(BATCH_SIZE)                                                #下一组数据集
             sess.run(train_op, feed_dict={x: xs, y_: ys})                                              #训练训练集数据
        test_acc = sess.run(accuracy, feed_dict=test_feed)                                          #测试准确率
        print("After %d training step(s), test accuracy using average""model is %g" % (TRAINING_STEPS, test_acc))

'''主函数：输入数据'''
def main(argv=None):
    mnist = input_data.read_data_sets("D:/python/pycharm/venv/tmp/data", one_hot=True)                  #输入训练数据
    train(mnist)

if __name__ == '__main__':
    tf.app.run()

三、最佳实践案例（变量管理及持久化的mnist实现）

1.def

       将神经网络的训练、测试、使用分成不同程序，并将前向传播网络抽象成一个独立的函数库，使整个流程更加灵活。

2.重要代码标注

（1）前向传播抽象成独立函数库

'''mnist_inference'''
import tensorflow as tf

INPUT_NODE = 784
OUTPUT_NODE = 10
LAYER1_NODE = 500

#通过tf.get_variable函数获取变量；在训练时创建变量，在测试时通过保存的模型加载这些变量的取值
def get_weight_variable(shape, regularizer):
    weights = tf.get_variable("weights", shape, initializer=tf.truncated_normal_initializer(stddev=0.1))
    if regularizer !=None:
        tf.add_to_collection('losses', regularizer(weights))
    return weights

#定义前向传播神经网络
def inference(input_tensor, regularizer):
    with tf.variable_scope('layer1'):                                                       #第一层神经网络
        weights = get_weight_variable([INPUT_NODE, LAYER1_NODE], regularizer)
        biases = tf.get_variable("biases", [LAYER1_NODE], initializer=tf.constant_initializer(0.0))
        layer1 = tf.nn.relu(tf.matmul(input_tensor, weights) + biases)
    with tf.variable_scope('layer2'):                                                       #第二层神经网络
        weights = get_weight_variable([LAYER1_NODE, OUTPUT_NODE], regularizer)
        biases = tf.get_variable("biases", [OUTPUT_NODE], initializer=tf.constant_initializer(0.0))
        layer2 = tf.matmul(layer1, weights) + biases
    return layer2

（2）训练：得出总损失

'''mnist_train.'''
import os
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import mnist_best

BATCH_SIZE = 100
LEARNING_RATE_BASE = 0.8
LEARNING_RATE_DECAY = 0.99
REGULARIZATION_RATE = 0.0001
TRAINING_STEPS = 30000
MOVING_AVERAGE_DECAY = 0.99
#模型保存的路径和文件名
MODEL_SAVE_PATH = "D:/python/pycharm/venv/best"
MODEL_NAME = "model.ckpt"

def train(mnist):
    x = tf.placeholder(tf.float32, [None, mnist_best.INPUT_NODE], name='x-input')     #调用mnist_best.py的INPUT 和 OUTPUT
    y_ = tf.placeholder(tf.float32, [None, mnist_best.OUTPUT_NODE], name='y-input')
    regularizer = tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)
    y = mnist_best.inference(x, regularizer)                                     #直接使用mnist_best.py定义的前向传播
    global_step = tf.Variable(0, trainable=False)
    variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
    variables_averages_op = variable_averages.apply(tf.trainable_variables())
    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=tf.argmax(y_, 1), logits=y)
    cross_entropy_mean = tf.reduce_mean(cross_entropy)
    loss = cross_entropy_mean + tf.add_n(tf.get_collection('losses'))
    learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE, global_step, mnist.train.num_examples / BATCH_SIZE, LEARNING_RATE_DECAY)
    train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
    with tf.control_dependencies([train_step, variables_averages_op]):
        train_op = tf.no_op(name='train')
    #初始化TensorFlow持久化类
    saver = tf.train.Saver()
    with tf.Session() as sess:
        tf.initialize_all_variables().run()
        for i in range(TRAINING_STEPS):
            xs, ys = mnist.train.next_batch(BATCH_SIZE)
            _, loss_value, step = sess.run([train_op, loss, global_step], feed_dict={x: xs, y_: ys})
            if i % 1000 == 0:
                 print("After %d training step,loss on training""batch is %g" % (step, loss_value))  #输出当前损失函数的大小
                 #保存当前模型
                 saver.save(sess, os.path.join(MODEL_SAVE_PATH, MODEL_NAME), global_step=global_step)

def main(argv=None):
    mnist = input_data.read_data_sets("D:/python/pycharm/venv/tmp/data", one_hot=True)
    train(mnist)

if __name__ == '__main__':
    tf.app.run()

（3）验证：验证准确率，每十秒测对应模型准确率一次

note：（2）（3）同时运行才能得到结果

'''测试程序 eval'''
import time
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import mnist_best
import mnist_best2

# 每10秒加载一次最新的模型，并且在测试数据上测试最新模型的正确率
EVAL_INTERVAL_SECS = 10

def evaluate(mnist):
    with tf.Graph().as_default() as g:
        x = tf.placeholder(tf.float32, [None, mnist_best.INPUT_NODE], name='x-input')    # 调用mnist_best.py的INPUT 和 OUTPUT
        y_ = tf.placeholder(tf.float32, [None, mnist_best.OUTPUT_NODE], name='y-input')
        validate_feed = {x: mnist.validation.images, y_: mnist.validation.labels}
        y = mnist_best.inference(x, None)                                                #测试时不关注ze正则化损失的值
        correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        variable_averages = tf.train.ExponentialMovingAverage(mnist_best2.MOVING_AVERAGE_DECAY)   # 通过变量重命名的方式来加载模型，
        #①生成字典 ②将滑动平均值赋值给v                                                         # 这样在前向传播的过程中就不需要调用求滑动平均的函数来获取平均值了。
        variables_to_restore = variable_averages.variables_to_restore()
        saver = tf.train.Saver(variables_to_restore)

        # 每隔EVAL_INTERVAL_SECS秒调用一次计算正确率的过程以检验训练过程中正确率的变化
        while True:
            with tf.Session() as sess:
                #找到最新建模文件名；tf.train.get_checkpoint_state函数会通过checkpoint文件自动找到目录中最新模型的文件名。
                ckpt = tf.train.get_checkpoint_state(mnist_best2.MODEL_SAVE_PATH)
                if ckpt and ckpt.model_checkpoint_path:
                    # 加载模型。
                    saver.restore(sess, ckpt.model_checkpoint_path)
                    # 通过文件名得到模型保存时迭代的轮数。
                    global_step = ckpt.model_checkpoint_path\
                                      .split('/')[-1].split('-')[-1]
                    accuracy_score = sess.run(accuracy, feed_dict=validate_feed)
                    print("After %s training step(s), validation ""accuracy = %g" % (global_step, accuracy_score))
                else:
                    print("No checkpoint file found")
                    return
            time.sleep(EVAL_INTERVAL_SECS)

def main(argv=None):
    mnist = input_data.read_data_sets("D:/python/pycharm/venv/tmp/data", one_hot=True)
    evaluate(mnist)

if __name__ == '__main__':
    tf.app.run()