深度神经网络中使用Tensorboard进行训练过程可视化

红太狼 2022-03-22 06:49 262阅读 0赞

**（转载，原文链接：[https://www.cnblogs.com/tengge/p/6376073.html）][https_www.cnblogs.com_tengge_p_6376073.html]**  
写得蛮不错的，亲测好用，关于怎么使用Tensorboard查看图片，请看这一篇的开头：[http://www.data-master.net/86658670][http_www.data-master.net_86658670]

import tensorflow as tf
    import numpy as np
    
    #输入数据
    x_data = np.linspace(-1,1,300)[:, np.newaxis]
    noise = np.random.normal(0,0.05, x_data.shape)
    y_data = np.square(x_data)-0.5+noise
    
    #输入层
    with tf.name_scope('input_layer'): #输入层。将这两个变量放到input_layer作用域下，tensorboard会把他们放在一个图形里面
        xs = tf.placeholder(tf.float32, [None, 1], name = 'x_input') # xs起名x_input，会在图形上显示
        ys = tf.placeholder(tf.float32, [None, 1], name = 'y_input') # ys起名y_input，会在图形上显示
    
    #隐层
    with tf.name_scope('hidden_layer'): #隐层。将隐层权重、偏置、净输入放在一起
        with tf.name_scope('weight'): #权重
            W1 = tf.Variable(tf.random_normal([1,10]))
            tf.summary.histogram('hidden_layer/weight', W1)
        with tf.name_scope('bias'): #偏置
            b1 = tf.Variable(tf.zeros([1,10])+0.1)
            tf.summary.histogram('hidden_layer/bias', b1)
        with tf.name_scope('Wx_plus_b'): #净输入
            Wx_plus_b1 = tf.matmul(xs,W1) + b1
            tf.summary.histogram('hidden_layer/Wx_plus_b',Wx_plus_b1)
    output1 = tf.nn.relu(Wx_plus_b1)
    
    #输出层
    with tf.name_scope('output_layer'): #输出层。将输出层权重、偏置、净输入放在一起
        with tf.name_scope('weight'): #权重
            W2 = tf.Variable(tf.random_normal([10,1]))
            tf.summary.histogram('output_layer/weight', W2)
        with tf.name_scope('bias'): #偏置
            b2 = tf.Variable(tf.zeros([1,1])+0.1)
            tf.summary.histogram('output_layer/bias', b2)
        with tf.name_scope('Wx_plus_b'): #净输入
            Wx_plus_b2 = tf.matmul(output1,W2) + b2
            tf.summary.histogram('output_layer/Wx_plus_b',Wx_plus_b2)
    output2 = Wx_plus_b2
    
    #损失
    with tf.name_scope('loss'): #损失
        loss = tf.reduce_mean(tf.reduce_sum(tf.square(ys-output2),reduction_indices=[1]))
        tf.summary.scalar('loss',loss)
    with tf.name_scope('train'): #训练过程
        train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)
    
    #初始化
    init = tf.global_variables_initializer()
    sess = tf.Session()
    sess.run(init)
    merged = tf.summary.merge_all() #将图形、训练过程等数据合并在一起
    writer = tf.summary.FileWriter('logs',sess.graph) #将训练日志写入到logs文件夹下
    
    #训练
    for i in range(1000):
        sess.run(train_step,feed_dict={xs:x_data,ys:y_data})
        if(i%50==0): #每50次写一次日志
            result = sess.run(merged,feed_dict={xs:x_data,ys:y_data}) #计算需要写入的日志数据
            writer.add_summary(result,i) #将日志数据写入文件

[https_www.cnblogs.com_tengge_p_6376073.html]: https://www.cnblogs.com/tengge/p/6376073.html%EF%BC%89
[http_www.data-master.net_86658670]: http://www.data-master.net/86658670