tensorboard 监控指标可视化

# -*- coding: utf-8 -*-

"""

Created on Wed Jan 30 23:24:30 2019

 

@author: Administrator

"""

 

import tensorflow as tf

from tensorflow.examples.tutorials.mnist import input_data

tf.reset_default_graph()

SUMMARY_DIR="log/"

BATCH_SIZE=100

TRAIN_STEPS=3000

#监控信息记入日志

#var：需要记账的张量

#name:图表名称

def variable_summaries(var,name):

    with tf.name_scope('summaries'):

        #记录张量的取值分布

        #记录到buffer中

        tf.summary.histogram(name,var)

        #求张量的所有元素的平均值

        mean=tf.reduce_mean(var)

        #日志

        #mean/为命名空间

        #相同命名空间下的监控指标会被整合到同一栏中。

        tf.summary.scalar('mean/'+name, mean)

        #标准差

        #stddev=tf.sqrt(tf.reduce_mean(tf.square(var-mean)))

        #tf.summary.scalar('stddev/'+name,var)

#生成一层全连接神经网络

def nn_layer(input_tensor,input_dim,output_dim,

             layer_name,

             act=tf.nn.relu):

    #将同一层神经网络放在同一命名空间下

    with tf.name_scope(layer_name):

        #声明神经网络边上的权重，并调用生成权重监控信息日志的函数

        with tf.name_scope("weights"):

            weights=tf.Variable(tf.truncated_normal(

                    [input_dim,output_dim],stddev=0.1))

            #权重记入日志

            variable_summaries(weights,layer_name+'/weights')

        #声明神经网络的偏置项

        with tf.name_scope('biases'):

            biases=tf.Variable(tf.constant(0.0,shape=[output_dim]))

            variable_summaries(biases,layer_name+'/biases')

        with tf.name_scope('Wx_plus_b'):

            preactivate=tf.matmul(input_tensor,weights)+biases

            #记录神经网络输出节点在经过激活函数之前的分布

            tf.summary.histogram(layer_name+'/pre_activate',

                                 preactivate)

        activations=act(preactivate,name='activation')

        #记录神经网络输出节点在经过激活函数之后的分布

        #

        tf.summary.histogram(layer_name+'/activations',activations)

        return activations

def main(argv=None):

    mnist=input_data.read_data_sets("MNIST_data/",one_hot=True)

    #定义输入

    with tf.name_scope("input"):

        x=tf.placeholder(

            tf.float32,[None,784],

            name='x-input')

        y_=tf.placeholder(

            tf.float32,[None,10],

            name='y-input')

       

    #将输入向量还原成图片的像素矩阵，并写入日志

    with tf.name_scope('input_reshape'):

        image_shaped_input=tf.reshape(x,[-1,28,28,1])

       

        tf.summary.image('input',image_shaped_input,10)

   

    hidden1=nn_layer(x,784,500,'layer1')

    y=nn_layer(hidden1,500,10,'layer2',act=tf.identity)

   

    #计算交叉熵，并记录日志

    with tf.name_scope('cross_entropy'):

        cross_entropy=tf.reduce_mean(

                tf.nn.softmax_cross_entropy_with_logits(

                labels=y_, logits=y)

                )

        tf.summary.scalar('cross_entropy',cross_entropy)

       

    with tf.name_scope('train'):

        train_step=tf.train.AdamOptimizer(0.001).minimize(cross_entropy)

   

    #计算模型在当前给定数据上的正确率

    #并日志

   

    with tf.name_scope('accuracy'):

        with tf.name_scope('correct_prediction'):

            correct_prediction=tf.equal(tf.argmax(y,1),tf.argmax(y_,1))

       

        with tf.name_scope('accuracy'):

            accuracy=tf.reduce_mean(

                    tf.cast(correct_prediction,tf.float32))

   

    #合并所有的日志操作

    merged=tf.summary.merge_all()

       

    with tf.Session() as sess:

        summary_writer=tf.summary.FileWriter('log/',sess.graph)

       

        tf.global_variables_initializer().run()

       

        for i in range(TRAIN_STEPS):

            xs,ys=mnist.train.next_batch(BATCH_SIZE)

           

            #调用训练

            summary,_=sess.run([merged,train_step],

                               feed_dict={x:xs, y_:ys})

           

            summary_writer.add_summary(summary,i)

        summary_writer.close()

 

if __name__=='__main__':

    tf.app.run()