深度學習-tensorflow基礎-變量作用域(命名空間)

Introduction

• 目的是為了使代碼更清晰區分功能，使其作用分明
• 讓tensorboard更清晰顯示代碼運作的模式

---

創建作用域

• 使用tf.variable_scope()函數創建作用域
  • scope_name參數：創建指定的名字
  • 其相當於是一種上下文環境
    • with tf.variable_scope("作用域名稱"):
    • 將相對應的操作放置到相對應的作用域中使graph更整齊

---

未使用作用域

以上一篇文章所使用的線性回歸代碼作為例子

import tensorflow as tf

def LinearRegression():
    X = tf.random_normal([100,1],mean=1.75,stddev=0.5, name="x_data")
    y_true = tf.matmul(X,[[0.7]]) + 0.8

    weight =  tf.Variable(tf.random_normal([1,1],mean=0.0,stddev=1.0),name="w") #必須用變量定義才能優化(改變)
    bias = tf.Variable(0.0,name="bias")

    y_predict = tf.matmul(X,weight) + bias

    loss = tf.reduce_mean(tf.square(y_true - y_predict))

    train_op = tf.train.GradientDescentOptimizer(0.1).minimize(loss)

    init_var_op = tf.global_variables_initializer()
    


    with tf.Session() as sess:
        sess.run(init_var_op)

        print("起始初始化權重：%f, 初始化偏置：%f"%(weight.eval(),bias.eval()))
        
        # 建立事件文件(指定存放文件夾及graph)
        FileWriter = tf.summary.FileWriter("./summary/",graph=sess.graph)

        for i in range(1000):
            sess.run(train_op)
            if i%50 == 0:
                print("優化%d次後 權重：%f, 優化偏置：%f" % (i,weight.eval(), bias.eval()))

if __name__ == '__main__':
    LinearRegression()

result

• 可以看到很多op裸露在外面較雜亂

使用作用域整理

import tensorflow as tf

def LinearRegression():
    with tf.variable_scope("data_preparation"):
        X = tf.random_normal([100,1],mean=1.75,stddev=0.5, name="x_data")
        y_true = tf.matmul(X,[[0.7]]) + 0.8

    with tf.variable_scope("LinearRegression_model_build"):
        weight =  tf.Variable(tf.random_normal([1,1],mean=0.0,stddev=1.0),name="w") #必須用變量定義才能優化(改變)
        bias = tf.Variable(0.0,name="bias")
        y_predict = tf.matmul(X,weight) + bias

    with tf.variable_scope("loss_calculate"):
        loss = tf.reduce_mean(tf.square(y_true - y_predict))

    with tf.variable_scope("optimize"):
        train_op = tf.train.GradientDescentOptimizer(0.1).minimize(loss)

    init_var_op = tf.global_variables_initializer()

    with tf.Session() as sess:
        sess.run(init_var_op)

        print("起始初始化權重：%f, 初始化偏置：%f"%(weight.eval(),bias.eval()))

        # 建立事件文件(指定存放文件夾及graph)
        FileWriter = tf.summary.FileWriter("./summary/",graph=sess.graph)

        for i in range(1000):
            sess.run(train_op)
            if i%50 == 0:
                print("優化%d次後 權重：%f, 優化偏置：%f" % (i,weight.eval(), bias.eval()))

if __name__ == '__main__':
    LinearRegression()

result

• 產生的graph比未整理過的清爽
• graph理解性變高，結構調理清楚