[吃药深度学习随笔] 学习率-CSDN博客

学习率（Learning_rate）

　　表示了每次参数更新的幅度大小

　　若过大则参数容易在最小值附近不断跳跃

　　若过小则参数收敛会变慢

于是如何设置学习率？

使用

指数衰减学习率

Tensorflow代码：

global_step = tf.Variable(0,trainable=False)
learning_rate = tf.train.exponential_decay(
    LEARNING_RATE_BASE,
    global_step,
    LEARNING_RATE_STEP,LEARNING_RATE_DECAY,
    staircase=True/False
)

例子：

import tensorflow as tf

LEARNING_RATE_BASE = 0.1 #最初学习率
LEARNING_RATE_DECAY = 0.99 #学习率衰减率
LEARNING_RATE_STEP = 1 #多少轮后更新学习率

#定义全局步数计数器
global_step = tf.Variable(0,trainable=False)
learning_rate = tf.train.exponential_decay(
    LEARNING_RATE_BASE,
    global_step,
    LEARNING_RATE_STEP,LEARNING_RATE_DECAY,
    staircase=True
)

#定义待优化参数
w = tf.Variable(tf.constant(5,dtype=tf.float32))
#定义损失函数loss
loss = tf.square(w+1)
#定义反向传播方法
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
#生成会话
with tf.Session() as sess:
    init_op = tf.global_variables_initializer()
    sess.run(init_op)
    for i in range(100):
        sess.run(train_step)
        learning_rate_val = sess.run(learning_rate)
        global_step_val = sess.run(global_step)
        w_val = sess.run(w)
        loss_val = sess.run(loss)
        print ("%s步后：global_step: %f , w: %f, Learning_rate = %f , loss = %f " %(i, global_step_val, w_val,
                                                                                  learning_rate_val, loss_val))

得到最终结果：

0步后：global_step: 1.000000 , w: 3.800000, Learning_rate = 0.099000 , loss = 23.040001 
1步后：global_step: 2.000000 , w: 2.849600, Learning_rate = 0.098010 , loss = 14.819419 
2步后：global_step: 3.000000 , w: 2.095001, Learning_rate = 0.097030 , loss = 9.579033 
3步后：global_step: 4.000000 , w: 1.494386, Learning_rate = 0.096060 , loss = 6.221961 
4步后：global_step: 5.000000 , w: 1.015167, Learning_rate = 0.095099 , loss = 4.060896 
5步后：global_step: 6.000000 , w: 0.631886, Learning_rate = 0.094148 , loss = 2.663051 
6步后：global_step: 7.000000 , w: 0.324608, Learning_rate = 0.093207 , loss = 1.754587 
7步后：global_step: 8.000000 , w: 0.077684, Learning_rate = 0.092274 , loss = 1.161403 
8步后：global_step: 9.000000 , w: -0.121202, Learning_rate = 0.091352 , loss = 0.772287 
......
90步后：global_step: 91.000000 , w: -0.999985, Learning_rate = 0.040069 , loss = 0.000000 
91步后：global_step: 92.000000 , w: -0.999987, Learning_rate = 0.039668 , loss = 0.000000 
92步后：global_step: 93.000000 , w: -0.999988, Learning_rate = 0.039271 , loss = 0.000000 
93步后：global_step: 94.000000 , w: -0.999989, Learning_rate = 0.038878 , loss = 0.000000 
94步后：global_step: 95.000000 , w: -0.999990, Learning_rate = 0.038490 , loss = 0.000000 
95步后：global_step: 96.000000 , w: -0.999990, Learning_rate = 0.038105 , loss = 0.000000 
96步后：global_step: 97.000000 , w: -0.999991, Learning_rate = 0.037724 , loss = 0.000000 
97步后：global_step: 98.000000 , w: -0.999992, Learning_rate = 0.037346 , loss = 0.000000 
98步后：global_step: 99.000000 , w: -0.999992, Learning_rate = 0.036973 , loss = 0.000000 
99步后：global_step: 100.000000 , w: -0.999993, Learning_rate = 0.036603 , loss = 0.000000