Tensorboard

tensorboard用以圖形化展示我們的代碼結構和圖形化訓練誤差等，輔助優化程序

參考鏈接：www.cnblogs.com/fydeblog/p/7429344.html

微代碼測試

代碼全文

import tensorflow as tf 

with tf.name_scope('graph') as scope:
    a = tf.constant([[3,4]],name = 'a')
    b = tf.constant([[5],[6]],name = 'b')
    product = tf.matmul(a,b,name='product')

sess = tf.Session()
writer = tf.summary.FileWriter('./tensorflow/',sess.graph)
init = tf.global_variables_initializer()
sess.run(init)
 

啟動tensorboard

運行代碼之後可以在目錄./tensorboard/ 下看到生成了一個文件，用於啟動tensorboard。執行tensorboard --logdir ./tensorflow/，打開瀏覽器127.0.0.1:6006 ,可以看到視圖

註：要保證該文件夾下只有一個文件

雙擊graph可以進一步打開，這就是我們這段微代碼的結構

進階

這裏我們基於之前的一段代碼，與tensorboard相結合來看下效果。
前文鏈接：https://www.cnblogs.com/maskerk/p/9973503.html

代碼全文

import tensorflow as tf 
import matplotlib.pyplot as plt
import numpy as np

#樣本數據
with tf.name_scope('sample-data'):
    x_train = np.linspace(-1,1,300)[:,np.newaxis]
    noise = np.random.normal(0, 0.1, x_train.shape)
    y_train = x_train * 3 + noise + 0.8

#
with tf.name_scope('hold-data'):
    x = tf.placeholder(tf.float32, [None, 1])
    y = tf.placeholder(tf.float32, [None, 1])

#線性模型
with tf.name_scope('line-model'):
    W = tf.Variable([0.1],dtype = tf.float32,name='W')
    #添加變量W到tensorboard的Distributions下
    tf.summary.histogram('Weight',W)
    b = tf.Variable([0.1],dtype = tf.float32,name='b')
    line_model = W * x + b
    #添加變量b到tensorboard的Distributions下
    tf.summary.histogram('bias',b)

#損失模型
with tf.name_scope("loss-model"):
    loss = tf.reduce_sum(tf.square(line_model - y))
    #添加變量loss到tensorboard的Scalars下
    tf.summary.scalar("loss",loss)
    

#創建優化器
with tf.name_scope("optimizer-model"):
    optimizer = tf.train.GradientDescentOptimizer(0.001)
    train = optimizer.minimize(loss)

#初始化變量
with tf.name_scope("init-model"):
    init = tf.global_variables_initializer()
    sess = tf.Session()
    sess.run(init)

# 繪制樣本數據
fig = plt.figure()
ax = fig.add_subplot(1,1,1)
ax.scatter(x_train, y_train)
plt.ion()
plt.show()
plt.pause(1)

#將所有的summary全部保存磁盤
merged = tf.summary.merge_all()

#tensorboard所需數據寫入文件
writer = tf.summary.FileWriter('./tensorflow/',sess.graph)

#訓練100次
for i in range(100):
    if i % 10 == 0:
        #每隔10次打印1次成果
        print(i)
        print('W:%s  b:%s' % (sess.run(W),sess.run(b)))
        print('loss:%s' % (sess.run(loss,{x:x_train,y:y_train})))

        #繪制擬合直線
        try:
            ax.lines.remove(lines[0])
        except Exception:
                        pass
        lines = ax.plot(x_train, sess.run(W)*x_train+sess.run(b), 'r-', lw=5)
        plt.pause(1)

    sess.run(train,{x:x_train,y:y_train})

    #向tensorboard添加數據
    rs = sess.run(merged,{x:x_train,y:y_train})
    writer.add_summary(rs,i)

# 打印訓練100次後的成果
print('---')
print('W:%s  b:%s' % (sess.run(W),sess.run(b)))
print('loss:%s' % (sess.run(loss,{x:x_train,y:y_train})))
 

相比前文，這裏添加了兩部分

• 1.擬合直線的動態變化圖像
  

• 2.tensorboard展現數據變化過程

啟動tensorboard

tensorboard --logdir ./tensorflow/

可以看到loss(誤差大小)的變化曲線

權重值W和偏差b的變化曲線(y = W * x + b)

tensorboard入門級測試