影醉阏轩窗

衣带渐宽终不悔,为伊消得人憔悴。
扩大
缩小

TessorFlow学习 之 神经网络的构建

 


   1.建立一个神经网络添加层

    输入值、输入的大小、输出的大小和激励函数

    学过神经网络的人看下面这个图就明白了,不懂的去看看我的另一篇博客

 

1 def add_layer(inputs , in_size , out_size , activate = None):
2     Weights = tf.Variable(tf.random_normal([in_size,out_size]))#随机初始化
3     baises  = tf.Variable(tf.zeros([1,out_size])+0.1)#可以随机但是不要初始化为0,都为固定值比随机好点
4     y = tf.matmul(inputs, Weights) + baises #matmul:矩阵乘法,multipy:一般是数量的乘法
5     if activate:
6         y = activate(y)
7     return y

   2.训练一个二次函数

 1 import tensorflow as tf
 2 import numpy as np
 3 
 4 def add_layer(inputs , in_size , out_size , activate = None):
 5     Weights = tf.Variable(tf.random_normal([in_size,out_size]))#随机初始化
 6     baises  = tf.Variable(tf.zeros([1,out_size])+0.1)#可以随机但是不要初始化为0,都为固定值比随机好点
 7     y = tf.matmul(inputs, Weights) + baises #matmul:矩阵乘法,multipy:一般是数量的乘法
 8     if activate:
 9         y = activate(y)
10     return y
11 if __name__ == '__main__':
12     x_data = np.linspace(-1,1,300,dtype=np.float32)[:,np.newaxis]#创建-1,1的300个数,此时为一维矩阵,后面转化为二维矩阵===[1,2,3]-->>[[1,2,3]]
13     noise = np.random.normal(0,0.05,x_data.shape).astype(np.float32)#噪声是(1,300)格式,0-0.05大小
14     y_data = np.square(x_data) - 0.5 + noise #带有噪声的抛物线
15 
16     xs = tf.placeholder(tf.float32,[None,1]) #外界输入数据
17     ys = tf.placeholder(tf.float32,[None,1])
18 
19     l1 = add_layer(xs,1,10,activate=tf.nn.relu)
20     prediction = add_layer(l1,10,1,activate=None)
21 
22     loss = tf.reduce_mean(tf.reduce_sum(tf.square(ys - prediction),reduction_indices=[1]))#误差
23     train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)#对误差进行梯度优化,步伐为0.1
24 
25     sess = tf.Session()
26     sess.run( tf.global_variables_initializer())
27     for i in range(1000):
28         sess.run(train_step, feed_dict={xs: x_data, ys: y_data})#训练
29         if i%50 == 0:
30             print(sess.run(loss, feed_dict={xs: x_data, ys: y_data}))#查看误差

   3.动态显示训练过程

    显示的步骤程序之中部分进行说明,其它说明请看其它博客

 1 import tensorflow as tf
 2 import numpy as np
 3 import matplotlib.pyplot as plt
 4 
 5 def add_layer(inputs , in_size , out_size , activate = None):
 6     Weights = tf.Variable(tf.random_normal([in_size,out_size]))#随机初始化
 7     baises  = tf.Variable(tf.zeros([1,out_size])+0.1)#可以随机但是不要初始化为0,都为固定值比随机好点
 8     y = tf.matmul(inputs, Weights) + baises #matmul:矩阵乘法,multipy:一般是数量的乘法
 9     if activate:
10         y = activate(y)
11     return y
12 if __name__ == '__main__':
13     x_data = np.linspace(-1,1,300,dtype=np.float32)[:,np.newaxis]#创建-1,1的300个数,此时为一维矩阵,后面转化为二维矩阵===[1,2,3]-->>[[1,2,3]]
14     noise = np.random.normal(0,0.05,x_data.shape).astype(np.float32)#噪声是(1,300)格式,0-0.05大小
15     y_data = np.square(x_data) - 0.5 + noise #带有噪声的抛物线
16     fig = plt.figure('show_data')# figure("data")指定图表名称
17     ax = fig.add_subplot(111)
18     ax.scatter(x_data,y_data)
19     plt.ion()
20     plt.show()
21     xs = tf.placeholder(tf.float32,[None,1]) #外界输入数据
22     ys = tf.placeholder(tf.float32,[None,1])
23 
24     l1 = add_layer(xs,1,10,activate=tf.nn.relu)
25     prediction = add_layer(l1,10,1,activate=None)
26 
27     loss = tf.reduce_mean(tf.reduce_sum(tf.square(ys - prediction),reduction_indices=[1]))#误差
28     train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)#对误差进行梯度优化,步伐为0.1
29 
30     sess = tf.Session()
31     sess.run( tf.global_variables_initializer())
32     for i in range(1000):
33         sess.run(train_step, feed_dict={xs: x_data, ys: y_data})#训练
34         if i%50 == 0:
35             try:
36                 ax.lines.remove(lines[0])
37             except Exception:
38                 pass
39             prediction_value = sess.run(prediction, feed_dict={xs: x_data})
40             lines = ax.plot(x_data,prediction_value,"r",lw = 3)
41             print(sess.run(loss, feed_dict={xs: x_data, ys: y_data}))#查看误差
42             plt.pause(2)
43     while True:
44         plt.pause(0.01)

 

 

   4.TensorBoard整体结构化显示

 

    A.利用with tf.name_scope("name")创建大结构、利用函数的name="name"去创建小结构:tf.placeholder(tf.float32,[None,1],name="x_data")

    B.利用writer = tf.summary.FileWriter("G:/test/",graph=sess.graph)创建一个graph文件

    C.利用TessorBoard去执行这个文件

      这里得注意--->>>首先到你存放文件的上一个目录--->>然后再去运行这个文件

      tensorboard  --logdir=test

   5.TensorBoard局部结构化显示  

      A. tf.summary.histogram(layer_name+"Weight",Weights):直方图显示

     

     B.  tf.summary.scalar("Loss",loss):折线图显示,loss的走向决定你的网络训练的好坏,至关重要一点

       C.初始化与运行设定的图表

1 merge = tf.summary.merge_all()#合并图表
2 writer = tf.summary.FileWriter("G:/test/",graph=sess.graph)#写进文件
3 result = sess.run(merge,feed_dict={xs:x_data,ys:y_data})#运行打包的图表merge
4 writer.add_summary(result,i)#写入文件,并且单步长50

     完整代码及显示效果:
 1 import tensorflow as tf
 2 import numpy as np
 3 import matplotlib.pyplot as plt
 4 
 5 def add_layer(inputs , in_size , out_size , n_layer = 1 , activate = None):
 6     layer_name = "layer" + str(n_layer)
 7     with tf.name_scope(layer_name):
 8         with tf.name_scope("Weights"):
 9             Weights = tf.Variable(tf.random_normal([in_size,out_size]),name="W")#随机初始化
10             tf.summary.histogram(layer_name+"Weight",Weights)
11         with tf.name_scope("Baises"):
12             baises  = tf.Variable(tf.zeros([1,out_size])+0.1,name="B")#可以随机但是不要初始化为0,都为固定值比随机好点
13             tf.summary.histogram(layer_name+"Baises",baises)
14         y = tf.matmul(inputs, Weights) + baises #matmul:矩阵乘法,multipy:一般是数量的乘法
15         if activate:
16             y = activate(y)
17         tf.summary.histogram(layer_name+"y_sum",y)
18         return y
19 if __name__ == '__main__':
20     x_data = np.linspace(-1,1,300,dtype=np.float32)[:,np.newaxis]#创建-1,1的300个数,此时为一维矩阵,后面转化为二维矩阵===[1,2,3]-->>[[1,2,3]]
21     noise = np.random.normal(0,0.05,x_data.shape).astype(np.float32)#噪声是(1,300)格式,0-0.05大小
22     y_data = np.square(x_data) - 0.5 + noise #带有噪声的抛物线
23     fig = plt.figure('show_data')# figure("data")指定图表名称
24     ax = fig.add_subplot(111)
25     ax.scatter(x_data,y_data)
26     plt.ion()
27     plt.show()
28     with tf.name_scope("inputs"):
29         xs = tf.placeholder(tf.float32,[None,1],name="x_data") #外界输入数据
30         ys = tf.placeholder(tf.float32,[None,1],name="y_data")
31     l1 = add_layer(xs,1,10,n_layer=1,activate=tf.nn.relu)
32     prediction = add_layer(l1,10,1,n_layer=2,activate=None)
33     with tf.name_scope("loss"):
34         loss = tf.reduce_mean(tf.reduce_sum(tf.square(ys - prediction),reduction_indices=[1]))#误差
35         tf.summary.scalar("Loss",loss)
36     with tf.name_scope("train_step"):
37         train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)#对误差进行梯度优化,步伐为0.1
38 
39     sess = tf.Session()
40     merge = tf.summary.merge_all()#合并
41     writer = tf.summary.FileWriter("G:/test/",graph=sess.graph)
42     sess.run( tf.global_variables_initializer())
43     for i in range(1000):
44         sess.run(train_step, feed_dict={xs: x_data, ys: y_data})#训练
45         if i%100 == 0:
46             result = sess.run(merge,feed_dict={xs:x_data,ys:y_data})#运行打包的图表merge
47             writer.add_summary(result,i)#写入文件,并且单步长50

注意: 假设你的py文件中写了tf的summary,并且存放在了此目录下“D:\test\logs” 调出cmd,cd到D:\test,然后输入tensorboard –logdir=logs。一定要cd到logs这个文件夹的上一级,其他会出现No graph definition files were found.问题。

 

 

主要参考莫凡大大:https://morvanzhou.github.io/

可视化出现问题了,参考这位大神:http://blog.csdn.net/fengying2016/article/details/54289931

posted on 2017-10-23 18:31  影醉阏轩窗  阅读(478)  评论(0编辑  收藏  举报

导航

/* 线条鼠标集合 */ /* 鼠标点击求赞文字特效 */ //带头像评论