TensorFlow学习笔记（七）Tesnor Board

　　为了更好的管理、调试和优化神经网络的训练过程，TensorFlow提供了一个可视化工具TensorBoard。TensorBoard可以有效的展示TensorFlow在运行过程中的计算图。、各种指标随着时间变化的趋势以及训练中使用到的腿昂等信息

　  一、TensorBoard简介

　　二、TensorBoard计算图可视化

　　　　1、命名空间与TensorBoard图上节点

　　　　2、节点信息

　　　　3、监控指标可视化

　  一、TensorBoard简介

　　　TensorBoard是 TensorFlow的可视化工具，它可以通过TensorFlow程序运行过程中输出的日志文件可视化TensorFlow的运行状态。TB与TF跑在不同分进程中。TB自动读取最新的TF日志文件，呈现当前TF的最新状态。

　　

import tensorflow as tf

#定义一个简单的计算图，实现向量的加法
input1 = tf.constant([1.0,2.0,3.0],name="input1")
input2 = tf.Variable(tf.random_uniform([3]),name="input2")
output = tf.add_n([input1,input2],name="output")
#生成一个写日志的writer，并将当前TF计算图写入日志
writer = tf.summary.FileWriter("path/to/log",graph=tf.get_default_graph())
writer.close()

通过命令tensorboard --logdir=path/to/log 来启动tensorboard

 

 

 

　　二、TensorBoard计算图可视化

　　　　1、命名空间与TensorBoard图上节点

　　　为了更好的组织可视化效果图上的计算节点，TB支持通过TF命名空间来整理可视化效果图上的节点。TensorFlow提供了两个命名空间函数tf.variable_scope和tf.name_scope。两者基本是等价的。唯一的区别是在使用tf.get_variable上有所不同。

　　

import tensorflow as tf

with tf.variable_scope("foo"):
    #在命名空间foo下，获取变量“bar”。得到变量 foo/bar
    a = tf.get_variable("bar",[1])
    print(a.name)

with tf.variable_scope("bar"):
    #在命名空间foo下，获取变量“bar”。得到变量 bar/bar.此时bar/bar和foo/bar并不冲突
    b = tf.get_variable("bar",[1])
    print(b.name)

with tf.name_scope("a"):
    #使用tf.Variable 会受到tf.name_scope影响。变量名为“b_1/Variable:0”
    a = tf.Variable([1])
    print(a.name)
    #使用tf.get_variable 不会受到tf.name_scope影响。变量名为“b:0”，没有加上name_scope的前缀
    b = tf.get_variable("b",[1])
    print(b.name)
with tf.name_scope("b"):
    #使用tf.Variable 会受到tf.name_scope影响。变量名为“b/Variable:0”
    a = tf.Variable([1])
    print(a.name)
    #使用tf.get_variable 不会受到tf.name_scope影响。变量名也为“b:0”，没有加上name_scope的前缀
    #会报错重复声明
    b = tf.get_variable("b",[1])
    print(b.name)

改进上一节的样例代码

import tensorflow as tf

with tf.name_scope("inout1"):
    input1 = tf.constant([1.0,2.0,3.0],name="input1")
with tf.name_scope("input2"):
    intput2 = tf.Variable(tf.random_uniform([3]),name="input2")
output = tf.add_n([input1,intput2],name="add")

writer = tf.summary.FileWriter('path/to/log',tf.get_default_graph())
writer.close()

可视化TensorFlow（五）中的样例程序

# -*- coding:utf-8 -*-
import os
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
#加载mnsit_inference.py中定义的变量和函数
from integerad_mnist import mnsit_inference1
import numpy as np

#配置神经网络的参数
BATCH_SIZE = 100
LR_BASE = 0.8
LR_DECAY = 0.99
REGULARAZTION_RATE = 0.0001
TRANING_STEPS = 30000
MOVING_AVERAGE_DECAY = 0.99
#模型保存的文件名和路径
MODEL_SAVE_PATH = "path/to/model/"
MODEL_SAVE_NAME = "model.ckpt"


INPUT_NODE = 784
OUTPUT_NODE =10
LAYER_NODE = 500

def train(mnsit):
    #定义输入和输出的placeholder,将处理输入数据的计算都放在“input”
    with tf.name_scope("input"):
        x = tf.placeholder(tf.float32,shape=[None,mnsit_inference1.INPUT_NODE],name="x_input")
        y_ = tf.placeholder(tf.float32,shape=[None,mnsit_inference1.OUTPUT_NODE],name="y_input")
    regularizer = tf.contrib.layers.l2_regularizer(REGULARAZTION_RATE)
    #直接使用mnsit_inference中定义的前向传播过程
    y = mnsit_inference1.inference(x,regularizer)
    global_step = tf.Variable(0,trainable=False)
    #将处理滑动平均相关的计算都放在moving_average命名空间下
    with tf.name_scope("moving_average"):
        variable_average = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY,global_step)
        variable_average_op = variable_average.apply(tf.trainable_variables())
    #将计算loss相关的计算都放在loss_func命名空间下
    with tf.name_scope("loss_func"):
        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"))
    #定义学习率、优化方法等放在“train_step”下
    with tf.name_scope("train_step"):
        learning_rate = tf.train.exponential_decay(LR_BASE,global_step,mnsit.train.num_examples/BATCH_SIZE,LR_DECAY)
        train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss,global_step=global_step)
        with tf.control_dependencies([train_step,variable_average_op]):
            train_op = tf.no_op("train")
    #初始化TF的持久化类
    saver = tf.train.Saver()
    with tf.Session() as sess:
        tf.initialize_all_variables().run()
        for i in range(TRANING_STEPS):
            xs,ys = mnsit.train.next_batch(BATCH_SIZE)
            _,loss_value,step = sess.run([train_op,loss,global_step],feed_dict={x:xs,y_:ys})
            #每1000轮保存一次模型
            if i % 1000 == 0:
                print("After {0} training steps,loss on training batch is {1}".format(step,loss_value))
                saver.save(sess,os.path.join(MODEL_SAVE_PATH,MODEL_SAVE_NAME),global_step=global_step)
    writer = tf.summary.FileWriter("path/to/log",tf.get_default_graph())
    writer.close()
def main(argv = None):
    mnsit = input_data.read_data_sets("mnist_set",one_hot=True)
    train(mnsit)
if __name__ == '__main__':
    tf.app.run()

生成的TB可视化

 

 　　　　除了手动的通过TensorFlow的命名空间来调整TensorBoard的可视化效果图，TensorFlow也会智能的调整可视化效果图上的节点。TB将TF分成了主图和辅助图。左侧的Graph为主图，右侧的Auxiliary Nodes为辅助图。TF会主动把连接表较多的点列出来放在辅助图中。

　　　　除了自动的方式，TF也支持手动的方式来调整可视化效果。

 

　　　　2、节点信息

　　除了展示TF计算图的结构，TB还可以展示TF计算图上每个节点的基本信息以及运行是所消耗的时间以及空间。

　　调整上面代码中迭代训练的部分，展示每次迭代TF计算节点运行时间和消耗的内存。

    with tf.Session() as sess:
        tf.initialize_all_variables().run()
        writer = tf.summary.FileWriter("path/to/log",tf.get_default_graph())
        for i in range(TRANING_STEPS):

            xs,ys = mnsit.train.next_batch(BATCH_SIZE)
            _,loss_value,step = sess.run([train_op,loss,global_step],feed_dict={x:xs,y_:ys})
            #每1000轮记录一次运行状态
            if i % 1000 == 0:
                #配置运行是需要记录的信息
                run_options =tf.RunOptions(trace_level = tf.RunOptions.FULL_TRACE)
                run_metadata = tf.RunMetadata()
                #将配置信息和记录运行是的元信息传入运行过程
                _,loss_value,step = sess.run([train_op,loss,global_step],feed_dict={x:xs,y_:ys},options=run_options,run_metadata=run_metadata)
                #将节点在运行是的信息写入日志
                writer.add_run_metadata(run_metadata,"step-%s"%i)
                print("After {0} training steps,loss on training batch is {1}".format(step,loss_value))
            else:
                 _,loss_value,step = sess.run([train_op,loss,global_step],feed_dict={x:xs,y_:ys})
    writer.close()

 

　　　　3、监控指标可视化

　　TB除了可视化TF的计算图，还可以可视化TF运行程序中各种有助于了解运行程序状态的监控指标。

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
SUMMARY_DIR = "path/to/log"
BATCH_SIZE =100
TRAIN_STEPS =30000

def variable_summaries(var,name):
    with tf.name_scope("summaries"):
        tf.summary.histogram(name,var)
        mean = tf.reduce_mean(var)
        tf.summary.scalar("mean/"+name,mean)
        stddev = tf.sqrt(tf.reduce_mean(tf.square(var-mean)))
        tf.summary.scalar("stddev/"+name,stddev)

#生成一层全连接层神经网络
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一个统一的空间
    with tf.name_scope(layer_name):
        with tf.name_scope("weights"):
            weights = tf.Variable(tf.truncated_normal([input_dim,output_dim],stddev=0.1))
            variable_summaries(weights,layer_name+'/weights')
        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
            tf.summary.histogram(layer_name+'/pre_activations',preactivate)
            activations = act(preactivate)
            tf.summary.histogram(layer_name+"/activations",activations)
            return activations
def main(_):
    mnsit = input_data.read_data_sets('mnist_set',one_hot=True)
    with tf.name_scope('input'):
        x = tf.placeholder(tf.float32,shape=[None,784],name='x_input')
        y_ = tf.placeholder(tf.float32,shape=[None,10],name='y_input')
    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')
    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_op = tf.train.AdamOptimizer(0.001).minimize(cross_entropy)
    with tf.name_scope('accuracy'):
        with tf.name_scope('correct_prediction'):
            correct_prediction = tf.equal(tf.arg_max(y,1),tf.argmax(y_,1))
        with tf.name_scope('accuracy'):
            accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
            tf.summary.scalar('accuracy',accuracy)
    merged = tf.summary.merge_all()

    with tf.Session() as sess :
        summary_writer = tf.summary.FileWriter(SUMMARY_DIR,sess.graph)
        tf.global_variables_initializer().run()
        for i in range(TRAIN_STEPS):
            xs,ys = mnsit.train.next_batch(BATCH_SIZE)
            summary,_ = sess.run([merged,train_op],feed_dict={x:xs,y_:ys})
            summary_writer.add_summary(summary,i)
    summary_writer.close()

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